add fcosr model (#6765)

* add fcosr

* fix some problem

* add docs for fcosr

* modify code

* modify focsr reader

* finish tensorrt deployment with dynamic shape

* modify according to review comment

Co-authored-by: wangxinxin08 <>

configs/rotate/README.md

@@ -16,6 +16,7 @@
 | 模型 | mAP | 学习率策略 | 角度表示 | 数据增广 | GPU数目 | 每GPU图片数目 | 模型下载 | 配置文件 |
 |:---:|:----:|:---------:|:-----:|:--------:|:-----:|:------------:|:-------:|:------:|
 | [S2ANet](./s2anet/README.md) | 73.84 | 2x | le135 | - | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/s2anet_alignconv_2x_dota.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/s2anet/s2anet_alignconv_2x_dota.yml) |
+| [FCOSR](./fcosr/README.md) | 76.62 | 3x | oc | - | 4 | 4 | [model](https://paddledet.bj.bcebos.com/models/fcosr_x50_3x_dota.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/fcosr/fcosr_x50_3x_dota.yml) |
 
 **注意:**
 

configs/rotate/README_en.md

@@ -15,6 +15,7 @@ Rotated object detection is used to detect rectangular bounding boxes with angle
 | Model | mAP | Lr Scheduler | Angle | Aug | GPU Number | images/GPU | download | config |
 |:---:|:----:|:---------:|:-----:|:--------:|:-----:|:------------:|:-------:|:------:|
 | [S2ANet](./s2anet/README_en.md) | 73.84 | 2x | le135 | - | 4 | 2 | [model](https://paddledet.bj.bcebos.com/models/s2anet_alignconv_2x_dota.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/s2anet/s2anet_alignconv_2x_dota.yml) |
+| [FCOSR](./fcosr/README_en.md) | 76.62 | 3x | oc | - | 4 | 4 | [model](https://paddledet.bj.bcebos.com/models/fcosr_x50_3x_dota.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/fcosr/fcosr_x50_3x_dota.yml) |
 
 **Notes:**
 

configs/rotate/fcosr/README.md

+简体中文 | [English](README_en.md)
+
+# FCOSR
+
+## 内容
+- [简介](#简介)
+- [模型库](#模型库)
+- [使用说明](#使用说明)
+- [预测部署](#预测部署)
+- [引用](#引用)
+
+## 简介
+
+[FCOSR](https://arxiv.org/abs/2111.10780)是基于[FCOS](https://arxiv.org/abs/1904.01355)的单阶段Anchor-Free的旋转框检测算法。FCOSR主要聚焦于旋转框的标签匹配策略，提出了椭圆中心采样和模糊样本标签匹配的方法。在loss方面，FCOSR使用了[ProbIoU](https://arxiv.org/abs/2106.06072)避免边界不连续性问题。
+
+## 模型库
+
+| 模型 | Backbone | mAP | 学习率策略 | 角度表示 | 数据增广 | GPU数目 | 每GPU图片数目 | 模型下载 | 配置文件 |
+|:---:|:--------:|:----:|:---------:|:-----:|:--------:|:-----:|:------------:|:-------:|:------:|
+| FCOSR-M | ResNeXt-50 | 76.62 | 3x | oc | - | 4 | 4 | [model](https://paddledet.bj.bcebos.com/models/fcosr_x50_3x_dota.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/fcosr/fcosr_x50_3x_dota.yml) |
+
+**注意:**
+
+- 如果**GPU卡数**或者**batch size**发生了改变，你需要按照公式 **lr<sub>new</sub> = lr<sub>default</sub> * (batch_size<sub>new</sub> * GPU_number<sub>new</sub>) / (batch_size<sub>default</sub> * GPU_number<sub>default</sub>)** 调整学习率。
+- 模型库中的模型默认使用单尺度训练单尺度测试。如果数据增广一栏标明MS，意味着使用多尺度训练和多尺度测试。如果数据增广一栏标明RR，意味着使用RandomRotate数据增广进行训练。
+
+## 使用说明
+
+参考[数据准备](../README.md#数据准备)准备数据。
+
+### 训练
+
+GPU单卡训练
+``` bash
+CUDA_VISIBLE_DEVICES=0 python tools/train.py -c configs/rotate/fcosr/fcosr_x50_3x_dota.yml
+```
+
+GPU多卡训练
+``` bash
+CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch --gpus 0,1,2,3 tools/train.py -c configs/rotate/fcosr/fcosr_x50_3x_dota.yml
+```
+
+### 预测
+
+执行以下命令预测单张图片，图片预测结果会默认保存在`output`文件夹下面
+``` bash
+python tools/infer.py -c configs/rotate/fcosr/fcosr_x50_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/fcosr_x50_3x_dota.pdparams --infer_img=demo/P0861__1.0__1154___824.png --draw_threshold=0.5
+```
+
+### DOTA数据集评估
+
+参考[DOTA Task](https://captain-whu.github.io/DOTA/tasks.html), 评估DOTA数据集需要生成一个包含所有检测结果的zip文件，每一类的检测结果储存在一个txt文件中，txt文件中每行格式为：`image_name score x1 y1 x2 y2 x3 y3 x4 y4`。将生成的zip文件提交到[DOTA Evaluation](https://captain-whu.github.io/DOTA/evaluation.html)的Task1进行评估。你可以执行以下命令得到test数据集的预测结果：
+``` bash
+python tools/infer.py -c configs/rotate/fcosr/fcosr_x50_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/fcosr_x50_3x_dota.pdparams --infer_dir=/path/to/test/images --output_dir=output_fcosr --visualize=False --save_results=True
+```
+将预测结果处理成官网评估所需要的格式：
+``` bash
+python configs/rotate/tools/generate_result.py --pred_txt_dir=output_fcosr/ --output_dir=submit/ --data_type=dota10
+
+zip -r submit.zip submit
+```
+
+## 预测部署
+
+部署教程请参考[预测部署](../../../deploy/README.md)
+
+## 引用
+
+```
+@article{li2021fcosr,
+  title={Fcosr: A simple anchor-free rotated detector for aerial object detection},
+  author={Li, Zhonghua and Hou, Biao and Wu, Zitong and Jiao, Licheng and Ren, Bo and Yang, Chen},
+  journal={arXiv preprint arXiv:2111.10780},
+  year={2021}
+}
+
+@inproceedings{tian2019fcos,
+  title={Fcos: Fully convolutional one-stage object detection},
+  author={Tian, Zhi and Shen, Chunhua and Chen, Hao and He, Tong},
+  booktitle={Proceedings of the IEEE/CVF international conference on computer vision},
+  pages={9627--9636},
+  year={2019}
+}
+
+@article{llerena2021gaussian,
+  title={Gaussian Bounding Boxes and Probabilistic Intersection-over-Union for Object Detection},
+  author={Llerena, Jeffri M and Zeni, Luis Felipe and Kristen, Lucas N and Jung, Claudio},
+  journal={arXiv preprint arXiv:2106.06072},
+  year={2021}
+}
+```

configs/rotate/fcosr/README_en.md

+English | [简体中文](README.md)
+
+# FCOSR
+
+## Content
+- [Introduction](#Introduction)
+- [Model Zoo](#Model-Zoo)
+- [Getting Start](#Getting-Start)
+- [Deployment](#Deployment)
+- [Citations](#Citations)
+
+## Introduction
+
+[FCOSR](https://arxiv.org/abs/2111.10780) is one stage anchor-free model based on [FCOS](https://arxiv.org/abs/1904.01355). FCOSR focuses on the label assignment strategy for oriented bounding boxes and proposes ellipse center sampling method and fuzzy sample assignment strategy. In terms of loss, FCOSR uses [ProbIoU](https://arxiv.org/abs/2106.06072) to avoid boundary discontinuity problem.
+
+## Model Zoo
+
+| Model | Backbone | mAP | Lr Scheduler | Angle | Aug | GPU Number | images/GPU | download | config |
+|:---:|:--------:|:----:|:---------:|:-----:|:--------:|:-----:|:------------:|:-------:|:------:|
+| FCOSR-M | ResNeXt-50 | 76.62 | 3x | oc | - | 4 | 4 | [model](https://paddledet.bj.bcebos.com/models/fcosr_x50_3x_dota.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/rotate/fcosr/fcosr_x50_3x_dota.yml) |
+
+**Notes:**
+
+- if **GPU number** or **mini-batch size** is changed, **learning rate** should be adjusted according to the formula **lr<sub>new</sub> = lr<sub>default</sub> * (batch_size<sub>new</sub> * GPU_number<sub>new</sub>) / (batch_size<sub>default</sub> * GPU_number<sub>default</sub>)**.
+- Models in model zoo is trained and tested with single scale by default. If `MS` is indicated in the data augmentation column, it means that multi-scale training and multi-scale testing are used. If `RR` is indicated in the data augmentation column, it means that RandomRotate data augmentation is used for training.
+
+## Getting Start
+
+Refer to [Data-Preparation](../README_en.md#Data-Preparation) to prepare data.
+
+### Training
+
+Single GPU Training
+``` bash
+CUDA_VISIBLE_DEVICES=0 python tools/train.py -c configs/rotate/fcosr/fcosr_x50_3x_dota.yml
+```
+
+Multiple GPUs Training
+``` bash
+CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch --gpus 0,1,2,3 tools/train.py -c configs/rotate/fcosr/fcosr_x50_3x_dota.yml
+```
+
+### Inference
+
+Run the follow command to infer single image, the result of inference will be saved in `output` directory by default.
+
+``` bash
+python tools/infer.py -c configs/rotate/fcosr/fcosr_x50_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/fcosr_x50_3x_dota.pdparams --infer_img=demo/P0861__1.0__1154___824.png --draw_threshold=0.5
+```
+
+### Evaluation on DOTA Dataset
+Refering to [DOTA Task](https://captain-whu.github.io/DOTA/tasks.html), You need to submit a zip file containing results for all test images for evaluation. The detection results of each category are stored in a txt file, each line of which is in the following format
+`image_id score x1 y1 x2 y2 x3 y3 x4 y4`. To evaluate, you should submit the generated zip file to the Task1 of [DOTA Evaluation](https://captain-whu.github.io/DOTA/evaluation.html). You can run the following command to get the inference results of test dataset:
+``` bash
+python tools/infer.py -c configs/rotate/fcosr/fcosr_x50_3x_dota.yml -o weights=https://paddledet.bj.bcebos.com/models/fcosr_x50_3x_dota.pdparams --infer_dir=/path/to/test/images --output_dir=output_fcosr --visualize=False --save_results=True
+```
+Process the prediction results into the format required for the official website evaluation:
+``` bash
+python configs/rotate/tools/generate_result.py --pred_txt_dir=output_fcosr/ --output_dir=submit/ --data_type=dota10
+
+zip -r submit.zip submit
+```
+
+## Deployment
+
+Please refer to the deployment tutorial[Deployment](../../../deploy/README_en.md)
+
+## Citations
+
+```
+@article{li2021fcosr,
+  title={Fcosr: A simple anchor-free rotated detector for aerial object detection},
+  author={Li, Zhonghua and Hou, Biao and Wu, Zitong and Jiao, Licheng and Ren, Bo and Yang, Chen},
+  journal={arXiv preprint arXiv:2111.10780},
+  year={2021}
+}
+
+@inproceedings{tian2019fcos,
+  title={Fcos: Fully convolutional one-stage object detection},
+  author={Tian, Zhi and Shen, Chunhua and Chen, Hao and He, Tong},
+  booktitle={Proceedings of the IEEE/CVF international conference on computer vision},
+  pages={9627--9636},
+  year={2019}
+}
+
+@article{llerena2021gaussian,
+  title={Gaussian Bounding Boxes and Probabilistic Intersection-over-Union for Object Detection},
+  author={Llerena, Jeffri M and Zeni, Luis Felipe and Kristen, Lucas N and Jung, Claudio},
+  journal={arXiv preprint arXiv:2106.06072},
+  year={2021}
+}
+```

configs/rotate/fcosr/_base_/fcosr_reader.yml

+worker_num: 4
+image_height: &image_height 1024
+image_width: &image_width 1024
+image_size: &image_size [*image_height, *image_width]
+
+TrainReader:
+  sample_transforms:
+    - Decode: {}
+    - Poly2Array: {}
+    - RandomRFlip: {}
+    - RandomRRotate: {angle_mode: 'value', angle: [0, 90, 180, -90]}
+    - RandomRRotate: {angle_mode: 'value', angle: [30, 60], rotate_prob: 0.5}
+    - RResize: {target_size: *image_size, keep_ratio: True, interp: 2}
+    - Poly2RBox: {filter_threshold: 2, filter_mode: 'edge', rbox_type: 'oc'}
+  batch_transforms:
+    - NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
+    - Permute: {}
+    - PadRGT: {}
+    - PadBatch: {pad_to_stride: 32}
+  batch_size: 4
+  shuffle: true
+  drop_last: true
+  use_shared_memory: true
+  collate_batch: true
+
+EvalReader:
+  sample_transforms:
+    - Decode: {}
+    - Poly2Array: {}
+    - RResize: {target_size: *image_size, keep_ratio: True, interp: 2}
+    - NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
+    - Permute: {}
+  batch_transforms:
+    - PadBatch: {pad_to_stride: 32}
+  batch_size: 2
+
+TestReader:
+  sample_transforms:
+    - Decode: {}
+    - Resize: {target_size: *image_size, keep_ratio: True, interp: 2}
+    - NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
+    - Permute: {}
+  batch_transforms:
+    - PadBatch: {pad_to_stride: 32}
+  batch_size: 8

configs/rotate/fcosr/_base_/fcosr_x50.yml

+architecture: YOLOv3
+snapshot_epoch: 1
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNeXt50_32x4d_pretrained.pdparams
+
+YOLOv3:
+  backbone: ResNet
+  neck: FPN
+  yolo_head: FCOSRHead
+  post_process: ~
+
+ResNet:
+  depth: 50
+  groups: 32
+  base_width: 4
+  variant: b
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [1,2,3]
+  num_stages: 4
+
+FPN:
+  out_channel: 256
+  extra_stage: 2
+  has_extra_convs: true
+  use_c5: false
+  relu_before_extra_convs: true
+
+FCOSRHead:
+  feat_channels: 256
+  fpn_strides: [8, 16, 32, 64, 128]
+  stacked_convs: 4
+  loss_weight: {class: 1.0, probiou: 1.0}
+  assigner:
+    name: FCOSRAssigner
+    factor: 12
+    threshold: 0.23
+    boundary: [[-1, 64], [64, 128], [128, 256], [256, 512], [512, 100000000.0]]
+  nms:
+    name: MultiClassNMS
+    nms_top_k: 2000
+    keep_top_k: -1
+    score_threshold: 0.1
+    nms_threshold: 0.1
+    normalized: False

configs/rotate/fcosr/_base_/optimizer_3x.yml

+epoch: 36
+
+LearningRate:
+  base_lr: 0.01
+  schedulers:
+    - !PiecewiseDecay
+      gamma: 0.1
+      milestones: [24, 33]
+    - !LinearWarmup
+      start_factor: 0.3333333
+      steps: 500
+
+OptimizerBuilder:
+  clip_grad_by_norm: 35.
+  optimizer:
+    momentum: 0.9
+    type: Momentum
+  regularizer:
+    factor: 0.0001
+    type: L2

configs/rotate/fcosr/fcosr_x50_3x_dota.yml

+_BASE_: [
+  '../../datasets/dota.yml',
+  '../../runtime.yml',
+  '_base_/optimizer_3x.yml',
+  '_base_/fcosr_reader.yml',
+  '_base_/fcosr_x50.yml'
+]
+
+weights: output/fcosr_x50_3x_dota/model_final

deploy/python/infer.py

@@ -856,13 +856,16 @@ def load_predictor(model_dir,
 
         if use_dynamic_shape:
             min_input_shape = {
-                'image': [batch_size, 3, trt_min_shape, trt_min_shape]
+                'image': [batch_size, 3, trt_min_shape, trt_min_shape],
+                'scale_factor': [batch_size, 2]
             }
             max_input_shape = {
-                'image': [batch_size, 3, trt_max_shape, trt_max_shape]
+                'image': [batch_size, 3, trt_max_shape, trt_max_shape],
+                'scale_factor': [batch_size, 2]
             }
             opt_input_shape = {
-                'image': [batch_size, 3, trt_opt_shape, trt_opt_shape]
+                'image': [batch_size, 3, trt_opt_shape, trt_opt_shape],
+                'scale_factor': [batch_size, 2]
             }
             config.set_trt_dynamic_shape_info(min_input_shape, max_input_shape,
                                               opt_input_shape)

ppdet/modeling/architectures/yolo.py

@@ -77,7 +77,10 @@ class YOLOv3(BaseArch):
 
     def _forward(self):
         body_feats = self.backbone(self.inputs)
+        if self.for_mot:
             neck_feats = self.neck(body_feats, self.for_mot)
+        else:
+            neck_feats = self.neck(body_feats)
 
         if isinstance(neck_feats, dict):
             assert self.for_mot == True

ppdet/modeling/assigners/__init__.py

@@ -17,9 +17,11 @@ from . import task_aligned_assigner
 from . import atss_assigner
 from . import simota_assigner
 from . import max_iou_assigner
+from . import fcosr_assigner
 
 from .utils import *
 from .task_aligned_assigner import *
 from .atss_assigner import *
 from .simota_assigner import *
 from .max_iou_assigner import *
+from .fcosr_assigner import *

ppdet/modeling/assigners/fcosr_assigner.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+
+from ppdet.core.workspace import register
+from ppdet.modeling.rbox_utils import box2corners, check_points_in_polys, paddle_gather
+
+__all__ = ['FCOSRAssigner']
+
+EPS = 1e-9
+
+
+@register
+class FCOSRAssigner(nn.Layer):
+    """ FCOSR Assigner, refer to https://arxiv.org/abs/2111.10780 for details
+
+    1. compute normalized gaussian distribution score and refined gaussian distribution score
+    2. refer to ellipse center sampling, sample points whose normalized gaussian distribution score is greater than threshold
+    3. refer to multi-level sampling, assign ground truth to feature map which follows two conditions.
+        i). first, the ratio between the short edge of the target and the stride of the feature map is less than 2.
+        ii). second, the long edge of minimum bounding rectangle of the target is larger than the acceptance range of feature map
+    4. refer to fuzzy sample label assignment, the points satisfying 2 and 3 will be assigned to the ground truth according to gaussian distribution score
+    """
+    __shared__ = ['num_classes']
+
+    def __init__(self,
+                 num_classes=80,
+                 factor=12,
+                 threshold=0.23,
+                 boundary=[[-1, 128], [128, 320], [320, 10000]],
+                 score_type='iou'):
+        super(FCOSRAssigner, self).__init__()
+        self.num_classes = num_classes
+        self.factor = factor
+        self.threshold = threshold
+        self.boundary = [
+            paddle.to_tensor(
+                l, dtype=paddle.float32).reshape([1, 1, 2]) for l in boundary
+        ]
+        self.score_type = score_type
+
+    def get_gaussian_distribution_score(self, points, gt_rboxes, gt_polys):
+        # projecting points to coordinate system defined by each rbox
+        # [B, N, 4, 2] -> 4 * [B, N, 1, 2]
+        a, b, c, d = gt_polys.split(4, axis=2)
+        # [1, L, 2] -> [1, 1, L, 2]
+        points = points.unsqueeze(0)
+        ab = b - a
+        ad = d - a
+        # [B, N, 5] -> [B, N, 2], [B, N, 2], [B, N, 1]
+        xy, wh, angle = gt_rboxes.split([2, 2, 1], axis=-1)
+        # [B, N, 2] -> [B, N, 1, 2]
+        xy = xy.unsqueeze(2)
+        # vector of points to center [B, N, L, 2]
+        vec = points - xy
+        # <ab, vec> = |ab| * |vec| * cos(theta) [B, N, L]
+        vec_dot_ab = paddle.sum(vec * ab, axis=-1)
+        # <ad, vec> = |ad| * |vec| * cos(theta) [B, N, L]
+        vec_dot_ad = paddle.sum(vec * ad, axis=-1)
+        # norm_ab [B, N, L]
+        norm_ab = paddle.sum(ab * ab, axis=-1).sqrt()
+        # norm_ad [B, N, L]
+        norm_ad = paddle.sum(ad * ad, axis=-1).sqrt()
+        # min(h, w), [B, N, 1]
+        min_edge = paddle.min(wh, axis=-1, keepdim=True)
+        # delta_x, delta_y [B, N, L]
+        delta_x = vec_dot_ab.pow(2) / (norm_ab.pow(3) * min_edge + EPS)
+        delta_y = vec_dot_ad.pow(2) / (norm_ad.pow(3) * min_edge + EPS)
+        # score [B, N, L]
+        norm_score = paddle.exp(-0.5 * self.factor * (delta_x + delta_y))
+
+        # simplified calculation
+        sigma = min_edge / self.factor
+        refined_score = norm_score / (2 * np.pi * sigma + EPS)
+        return norm_score, refined_score
+
+    def get_rotated_inside_mask(self, points, gt_polys, scores):
+        inside_mask = check_points_in_polys(points, gt_polys)
+        center_mask = scores >= self.threshold
+        return (inside_mask & center_mask).cast(paddle.float32)
+
+    def get_inside_range_mask(self, points, gt_bboxes, gt_rboxes, stride_tensor,
+                              regress_range):
+        # [1, L, 2] -> [1, 1, L, 2]
+        points = points.unsqueeze(0)
+        # [B, n, 4] -> [B, n, 1, 4]
+        x1y1, x2y2 = gt_bboxes.unsqueeze(2).split(2, axis=-1)
+        # [B, n, L, 2]
+        lt = points - x1y1
+        rb = x2y2 - points
+        # [B, n, L, 4]
+        ltrb = paddle.concat([lt, rb], axis=-1)
+        # [B, n, L, 4] -> [B, n, L]
+        inside_mask = paddle.min(ltrb, axis=-1) > EPS
+        # regress_range [1, L, 2] -> [1, 1, L, 2]
+        regress_range = regress_range.unsqueeze(0)
+        # stride_tensor [1, L, 1] -> [1, 1, L]
+        stride_tensor = stride_tensor.transpose((0, 2, 1))
+        # fcos range
+        # [B, n, L, 4] -> [B, n, L]
+        ltrb_max = paddle.max(ltrb, axis=-1)
+        # [1, 1, L, 2] -> [1, 1, L]
+        low, high = regress_range[..., 0], regress_range[..., 1]
+        # [B, n, L]
+        regress_mask = (ltrb_max >= low) & (ltrb_max <= high)
+        # mask for rotated
+        # [B, n, 1]
+        min_edge = paddle.min(gt_rboxes[..., 2:4], axis=-1, keepdim=True)
+        # [B, n , L]
+        rotated_mask = ((min_edge / stride_tensor) < 2.0) & (ltrb_max > high)
+        mask = inside_mask & (regress_mask | rotated_mask)
+        return mask.cast(paddle.float32)
+
+    @paddle.no_grad()
+    def forward(self,
+                anchor_points,
+                stride_tensor,
+                num_anchors_list,
+                gt_labels,
+                gt_bboxes,
+                gt_rboxes,
+                pad_gt_mask,
+                bg_index,
+                pred_rboxes=None):
+        r"""
+
+        Args:
+            anchor_points (Tensor, float32): pre-defined anchor points, shape(1, L, 2),
+                    "x, y" format
+            stride_tensor (Tensor, float32): stride tensor, shape (1, L, 1)
+            num_anchors_list (List): num of anchors in each level
+            gt_labels (Tensor, int64|int32): Label of gt_bboxes, shape(B, n, 1)
+            gt_bboxes (Tensor, float32): Ground truth bboxes, shape(B, n, 4)
+            gt_rboxes (Tensor, float32): Ground truth bboxes, shape(B, n, 5)
+            pad_gt_mask (Tensor, float32): 1 means bbox, 0 means no bbox, shape(B, n, 1)
+            bg_index (int): background index
+            pred_rboxes (Tensor, float32, optional): predicted bounding boxes, shape(B, L, 5)
+        Returns:
+            assigned_labels (Tensor): (B, L)
+            assigned_rboxes (Tensor): (B, L, 5)
+            assigned_scores (Tensor): (B, L, C), if pred_rboxes is not None, then output ious
+        """
+
+        _, num_anchors, _ = anchor_points.shape
+        batch_size, num_max_boxes, _ = gt_rboxes.shape
+        if num_max_boxes == 0:
+            assigned_labels = paddle.full(
+                [batch_size, num_anchors], bg_index, dtype=gt_labels.dtype)
+            assigned_rboxes = paddle.zeros([batch_size, num_anchors, 5])
+            assigned_scores = paddle.zeros(
+                [batch_size, num_anchors, self.num_classes])
+            return assigned_labels, assigned_rboxes, assigned_scores
+
+        # get normalized gaussian distribution score and refined distribution score
+        gt_polys = box2corners(gt_rboxes)
+        score, refined_score = self.get_gaussian_distribution_score(
+            anchor_points, gt_rboxes, gt_polys)
+        inside_mask = self.get_rotated_inside_mask(anchor_points, gt_polys,
+                                                   score)
+        regress_ranges = []
+        for num, bound in zip(num_anchors_list, self.boundary):
+            regress_ranges.append(bound.tile((1, num, 1)))
+        regress_ranges = paddle.concat(regress_ranges, axis=1)
+        regress_mask = self.get_inside_range_mask(
+            anchor_points, gt_bboxes, gt_rboxes, stride_tensor, regress_ranges)
+        # [B, n, L]
+        mask_positive = inside_mask * regress_mask * pad_gt_mask
+        refined_score = refined_score * mask_positive - (1. - mask_positive)
+
+        argmax_refined_score = refined_score.argmax(axis=-2)
+        max_refined_score = refined_score.max(axis=-2)
+        assigned_gt_index = argmax_refined_score
+
+        # assigned target
+        batch_ind = paddle.arange(
+            end=batch_size, dtype=gt_labels.dtype).unsqueeze(-1)
+        assigned_gt_index = assigned_gt_index + batch_ind * num_max_boxes
+        assigned_labels = paddle.gather(
+            gt_labels.flatten(), assigned_gt_index.flatten(), axis=0)
+        assigned_labels = assigned_labels.reshape([batch_size, num_anchors])
+        assigned_labels = paddle.where(
+            max_refined_score > 0, assigned_labels,
+            paddle.full_like(assigned_labels, bg_index))
+
+        assigned_rboxes = paddle.gather(
+            gt_rboxes.reshape([-1, 5]), assigned_gt_index.flatten(), axis=0)
+        assigned_rboxes = assigned_rboxes.reshape([batch_size, num_anchors, 5])
+
+        assigned_scores = F.one_hot(assigned_labels, self.num_classes + 1)
+        ind = list(range(self.num_classes + 1))
+        ind.remove(bg_index)
+        assigned_scores = paddle.index_select(
+            assigned_scores, paddle.to_tensor(ind), axis=-1)
+
+        if self.score_type == 'gaussian':
+            selected_scores = paddle_gather(
+                score, 1, argmax_refined_score.unsqueeze(-2)).squeeze(-2)
+            assigned_scores = assigned_scores * selected_scores.unsqueeze(-1)
+        elif self.score_type == 'iou':
+            assert pred_rboxes is not None, 'If score type is iou, pred_rboxes should not be None'
+            from ext_op import matched_rbox_iou
+            b, l = pred_rboxes.shape[:2]
+            iou_score = matched_rbox_iou(
+                pred_rboxes.reshape((-1, 5)), assigned_rboxes.reshape(
+                    (-1, 5))).reshape((b, l, 1))
+            assigned_scores = assigned_scores * iou_score
+
+        return assigned_labels, assigned_rboxes, assigned_scores
\ No newline at end of file

ppdet/modeling/heads/__init__.py

@@ -33,6 +33,7 @@ from . import sparsercnn_head
 from . import tood_head
 from . import retina_head
 from . import ppyoloe_head
+from . import fcosr_head
 
 from .bbox_head import *
 from .mask_head import *
@@ -55,3 +56,4 @@ from .sparsercnn_head import *
 from .tood_head import *
 from .retina_head import *
 from .ppyoloe_head import *
+from .fcosr_head import *

ppdet/modeling/heads/fcosr_head.py

+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+from ppdet.core.workspace import register
+from paddle import ParamAttr
+from paddle.regularizer import L2Decay
+
+from .fcos_head import ScaleReg
+from ..initializer import bias_init_with_prob, constant_, normal_
+from ..ops import get_act_fn, anchor_generator
+from ..rbox_utils import box2corners
+from ..losses import ProbIoULoss
+import numpy as np
+
+__all__ = ['FCOSRHead']
+
+
+def trunc_div(a, b):
+    ipt = paddle.divide(a, b)
+    sign_ipt = paddle.sign(ipt)
+    abs_ipt = paddle.abs(ipt)
+    abs_ipt = paddle.floor(abs_ipt)
+    out = paddle.multiply(sign_ipt, abs_ipt)
+    return out
+
+
+def fmod(a, b):
+    return a - trunc_div(a, b) * b
+
+
+def fmod_eval(a, b):
+    return a - a.divide(b).cast(paddle.int32).cast(paddle.float32) * b
+
+
+class ConvBNLayer(nn.Layer):
+    def __init__(self,
+                 ch_in,
+                 ch_out,
+                 filter_size=3,
+                 stride=1,
+                 groups=1,
+                 padding=0,
+                 norm_cfg={'name': 'gn',
+                           'num_groups': 32},
+                 act=None):
+        super(ConvBNLayer, self).__init__()
+
+        self.conv = nn.Conv2D(
+            in_channels=ch_in,
+            out_channels=ch_out,
+            kernel_size=filter_size,
+            stride=stride,
+            padding=padding,
+            groups=groups,
+            bias_attr=False)
+
+        norm_type = norm_cfg['name']
+        if norm_type in ['sync_bn', 'bn']:
+            self.norm = nn.BatchNorm2D(
+                ch_out,
+                weight_attr=ParamAttr(regularizer=L2Decay(0.0)),
+                bias_attr=ParamAttr(regularizer=L2Decay(0.0)))
+        else:
+            groups = norm_cfg.get('num_groups', 1)
+            self.norm = nn.GroupNorm(
+                num_groups=groups,
+                num_channels=ch_out,
+                weight_attr=ParamAttr(regularizer=L2Decay(0.0)),
+                bias_attr=ParamAttr(regularizer=L2Decay(0.0)))
+        self.act = get_act_fn(act) if act is None or isinstance(act, (
+            str, dict)) else act
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.norm(x)
+        x = self.act(x)
+
+        return x
+
+
+@register
+class FCOSRHead(nn.Layer):
+    """ FCOSR Head, refer to https://arxiv.org/abs/2111.10780 for details """
+
+    __shared__ = ['num_classes', 'trt']
+    __inject__ = ['assigner', 'nms']
+
+    def __init__(self,
+                 num_classes=15,
+                 in_channels=256,
+                 feat_channels=256,
+                 stacked_convs=4,
+                 act='relu',
+                 fpn_strides=[4, 8, 16, 32, 64],
+                 trt=False,
+                 loss_weight={'class': 1.0,
+                              'probiou': 1.0},
+                 norm_cfg={'name': 'gn',
+                           'num_groups': 32},
+                 assigner='FCOSRAssigner',
+                 nms='MultiClassNMS'):
+
+        super(FCOSRHead, self).__init__()
+        self.in_channels = in_channels
+        self.num_classes = num_classes
+        self.fpn_strides = fpn_strides
+        self.stacked_convs = stacked_convs
+        self.loss_weight = loss_weight
+        self.half_pi = paddle.to_tensor(
+            [1.5707963267948966], dtype=paddle.float32)
+        self.probiou_loss = ProbIoULoss(mode='l1')
+        act = get_act_fn(
+            act, trt=trt) if act is None or isinstance(act,
+                                                       (str, dict)) else act
+        self.trt = trt
+        self.loss_weight = loss_weight
+        self.assigner = assigner
+        self.nms = nms
+        # stem
+        self.stem_cls = nn.LayerList()
+        self.stem_reg = nn.LayerList()
+        for i in range(self.stacked_convs):
+            self.stem_cls.append(
+                ConvBNLayer(
+                    self.in_channels[i],
+                    feat_channels,
+                    filter_size=3,
+                    stride=1,
+                    padding=1,
+                    norm_cfg=norm_cfg,
+                    act=act))
+            self.stem_reg.append(
+                ConvBNLayer(
+                    self.in_channels[i],
+                    feat_channels,
+                    filter_size=3,
+                    stride=1,
+                    padding=1,
+                    norm_cfg=norm_cfg,
+                    act=act))
+
+        self.scales = nn.LayerList(
+            [ScaleReg() for _ in range(len(fpn_strides))])
+
+        # prediction
+        self.pred_cls = nn.Conv2D(feat_channels, self.num_classes, 3, padding=1)
+
+        self.pred_xy = nn.Conv2D(feat_channels, 2, 3, padding=1)
+
+        self.pred_wh = nn.Conv2D(feat_channels, 2, 3, padding=1)
+
+        self.pred_angle = nn.Conv2D(feat_channels, 1, 3, padding=1)
+
+        self._init_weights()
+
+    def _init_weights(self):
+        for cls_, reg_ in zip(self.stem_cls, self.stem_reg):
+            normal_(cls_.conv.weight, std=0.01)
+            normal_(reg_.conv.weight, std=0.01)
+
+        bias_cls = bias_init_with_prob(0.01)
+        normal_(self.pred_cls.weight, std=0.01)
+        constant_(self.pred_cls.bias, bias_cls)
+        normal_(self.pred_xy.weight, std=0.01)
+        normal_(self.pred_wh.weight, std=0.01)
+        normal_(self.pred_angle.weight, std=0.01)
+
+    @classmethod
+    def from_config(cls, cfg, input_shape):
+        return {'in_channels': [i.channels for i in input_shape], }
+
+    def _generate_anchors(self, feats):
+        if self.trt:
+            anchor_points = []
+            for feat, stride in zip(feats, self.fpn_strides):
+                _, _, h, w = paddle.shape(feat)
+                anchor, _ = anchor_generator(
+                    feat,
+                    stride * 4,
+                    1.0, [1.0, 1.0, 1.0, 1.0], [stride, stride],
+                    offset=0.5)
+                x1, y1, x2, y2 = paddle.split(anchor, 4, axis=-1)
+                xc = (x1 + x2 + 1) / 2
+                yc = (y1 + y2 + 1) / 2
+                anchor_point = paddle.concat(
+                    [xc, yc], axis=-1).reshape((1, h * w, 2))
+                anchor_points.append(anchor_point)
+            anchor_points = paddle.concat(anchor_points, axis=1)
+            return anchor_points, None, None
+        else:
+            anchor_points = []
+            stride_tensor = []
+            num_anchors_list = []
+            for i, stride in enumerate(self.fpn_strides):
+                _, _, h, w = feats[i].shape
+                shift_x = (paddle.arange(end=w) + 0.5) * stride
+                shift_y = (paddle.arange(end=h) + 0.5) * stride
+                shift_y, shift_x = paddle.meshgrid(shift_y, shift_x)
+                anchor_point = paddle.cast(
+                    paddle.stack(
+                        [shift_x, shift_y], axis=-1), dtype='float32')
+                anchor_points.append(anchor_point.reshape([1, -1, 2]))
+                stride_tensor.append(
+                    paddle.full(
+                        [1, h * w, 1], stride, dtype='float32'))
+                num_anchors_list.append(h * w)
+            anchor_points = paddle.concat(anchor_points, axis=1)
+            stride_tensor = paddle.concat(stride_tensor, axis=1)
+            return anchor_points, stride_tensor, num_anchors_list
+
+    def forward(self, feats, target=None):
+        if self.training:
+            return self.forward_train(feats, target)
+        else:
+            return self.forward_eval(feats, target)
+
+    def forward_train(self, feats, target=None):
+        anchor_points, stride_tensor, num_anchors_list = self._generate_anchors(
+            feats)
+        cls_pred_list, reg_pred_list = [], []
+        for stride, feat, scale in zip(self.fpn_strides, feats, self.scales):
+            # cls
+            cls_feat = feat
+            for cls_layer in self.stem_cls:
+                cls_feat = cls_layer(cls_feat)
+            cls_pred = F.sigmoid(self.pred_cls(cls_feat))
+            cls_pred_list.append(cls_pred.flatten(2).transpose((0, 2, 1)))
+            # reg
+            reg_feat = feat
+            for reg_layer in self.stem_reg:
+                reg_feat = reg_layer(reg_feat)
+
+            reg_xy = scale(self.pred_xy(reg_feat)) * stride
+            reg_wh = F.elu(scale(self.pred_wh(reg_feat)) + 1.) * stride
+            reg_angle = self.pred_angle(reg_feat)
+            reg_angle = fmod(reg_angle, self.half_pi)
+            reg_pred = paddle.concat([reg_xy, reg_wh, reg_angle], axis=1)
+            reg_pred_list.append(reg_pred.flatten(2).transpose((0, 2, 1)))
+
+        cls_pred_list = paddle.concat(cls_pred_list, axis=1)
+        reg_pred_list = paddle.concat(reg_pred_list, axis=1)
+
+        return self.get_loss([
+            cls_pred_list, reg_pred_list, anchor_points, stride_tensor,
+            num_anchors_list
+        ], target)
+
+    def forward_eval(self, feats, target=None):
+        cls_pred_list, reg_pred_list = [], []
+        anchor_points, _, _ = self._generate_anchors(feats)
+        for stride, feat, scale in zip(self.fpn_strides, feats, self.scales):
+            b, _, h, w = paddle.shape(feat)
+            # cls
+            cls_feat = feat
+            for cls_layer in self.stem_cls:
+                cls_feat = cls_layer(cls_feat)
+            cls_pred = F.sigmoid(self.pred_cls(cls_feat))
+            cls_pred_list.append(cls_pred.reshape([b, self.num_classes, h * w]))
+            # reg
+            reg_feat = feat
+            for reg_layer in self.stem_reg:
+                reg_feat = reg_layer(reg_feat)
+
+            reg_xy = scale(self.pred_xy(reg_feat)) * stride
+            reg_wh = F.elu(scale(self.pred_wh(reg_feat)) + 1.) * stride
+            reg_angle = self.pred_angle(reg_feat)
+            reg_angle = fmod_eval(reg_angle, self.half_pi)
+            reg_pred = paddle.concat([reg_xy, reg_wh, reg_angle], axis=1)
+            reg_pred = reg_pred.reshape([b, 5, h * w]).transpose((0, 2, 1))
+            reg_pred_list.append(reg_pred)
+
+        cls_pred_list = paddle.concat(cls_pred_list, axis=2)
+        reg_pred_list = paddle.concat(reg_pred_list, axis=1)
+        reg_pred_list = self._bbox_decode(anchor_points, reg_pred_list)
+        return cls_pred_list, reg_pred_list
+
+    def _bbox_decode(self, points, reg_pred_list):
+        xy, wha = paddle.split(reg_pred_list, [2, 3], axis=-1)
+        xy = xy + points
+        return paddle.concat([xy, wha], axis=-1)
+
+    def _box2corners(self, pred_bboxes):
+        """ convert (x, y, w, h, angle) to (x1, y1, x2, y2, x3, y3, x4, y4)
+
+        Args:
+            pred_bboxes (Tensor): [B, N, 5]
+        
+        Returns:
+            polys (Tensor): [B, N, 8]
+        """
+        x, y, w, h, angle = paddle.split(pred_bboxes, 5, axis=-1)
+        cos_a_half = paddle.cos(angle) * 0.5
+        sin_a_half = paddle.sin(angle) * 0.5
+        w_x = cos_a_half * w
+        w_y = sin_a_half * w
+        h_x = -sin_a_half * h
+        h_y = cos_a_half * h
+        return paddle.concat(
+            [
+                x + w_x + h_x, y + w_y + h_y, x - w_x + h_x, y - w_y + h_y,
+                x - w_x - h_x, y - w_y - h_y, x + w_x - h_x, y + w_y - h_y
+            ],
+            axis=-1)
+
+    def get_loss(self, head_outs, gt_meta):
+        cls_pred_list, reg_pred_list, anchor_points, stride_tensor, num_anchors_list = head_outs
+        gt_labels = gt_meta['gt_class']
+        gt_bboxes = gt_meta['gt_bbox']
+        gt_rboxes = gt_meta['gt_rbox']
+        pad_gt_mask = gt_meta['pad_gt_mask']
+        # decode
+        pred_rboxes = self._bbox_decode(anchor_points, reg_pred_list)
+        # label assignment
+        assigned_labels, assigned_rboxes, assigned_scores = \
+            self.assigner(
+                anchor_points,
+                stride_tensor,
+                num_anchors_list,
+                gt_labels,
+                gt_bboxes,
+                gt_rboxes,
+                pad_gt_mask,
+                self.num_classes,
+                pred_rboxes
+            )
+
+        # reg_loss
+        mask_positive = (assigned_labels != self.num_classes)
+        num_pos = mask_positive.sum().item()
+        if num_pos > 0:
+            bbox_mask = mask_positive.unsqueeze(-1).tile([1, 1, 5])
+            pred_rboxes_pos = paddle.masked_select(pred_rboxes,
+                                                   bbox_mask).reshape([-1, 5])
+            assigned_rboxes_pos = paddle.masked_select(
+                assigned_rboxes, bbox_mask).reshape([-1, 5])
+            bbox_weight = paddle.masked_select(
+                assigned_scores.sum(-1), mask_positive).reshape([-1])
+            avg_factor = bbox_weight.sum()
+            loss_probiou = self.probiou_loss(pred_rboxes_pos,
+                                             assigned_rboxes_pos)
+            loss_probiou = paddle.sum(loss_probiou * bbox_weight) / avg_factor
+        else:
+            loss_probiou = pred_rboxes.sum() * 0.
+
+        avg_factor = max(num_pos, 1.0)
+        # cls_loss
+        loss_cls = self._qfocal_loss(
+            cls_pred_list, assigned_scores, reduction='sum')
+        loss_cls = loss_cls / avg_factor
+
+        loss = self.loss_weight['class'] * loss_cls + \
+               self.loss_weight['probiou'] * loss_probiou
+        out_dict = {
+            'loss': loss,
+            'loss_probiou': loss_probiou,
+            'loss_cls': loss_cls
+        }
+        return out_dict
+
+    @staticmethod
+    def _qfocal_loss(score, label, gamma=2.0, reduction='sum'):
+        weight = (score - label).pow(gamma)
+        loss = F.binary_cross_entropy(
+            score, label, weight=weight, reduction=reduction)
+        return loss
+
+    def post_process(self, head_outs, scale_factor):
+        pred_scores, pred_rboxes = head_outs
+        # [B, N, 5] -> [B, N, 4, 2] -> [B, N, 8]
+        pred_rboxes = self._box2corners(pred_rboxes)
+        # scale bbox to origin
+        scale_y, scale_x = paddle.split(scale_factor, 2, axis=-1)
+        scale_factor = paddle.concat(
+            [
+                scale_x, scale_y, scale_x, scale_y, scale_x, scale_y, scale_x,
+                scale_y
+            ],
+            axis=-1).reshape([-1, 1, 8])
+        pred_rboxes /= scale_factor
+        bbox_pred, bbox_num, _ = self.nms(pred_rboxes, pred_scores)
+        return bbox_pred, bbox_num

ppdet/modeling/losses/__init__.py

@@ -27,6 +27,7 @@ from . import detr_loss
 from . import sparsercnn_loss
 from . import focal_loss
 from . import smooth_l1_loss
+from . import probiou_loss
 
 from .yolo_loss import *
 from .iou_aware_loss import *
@@ -44,3 +45,4 @@ from .sparsercnn_loss import *
 from .focal_loss import *
 from .smooth_l1_loss import *
 from .pose3d_loss import *
+from .probiou_loss import *

ppdet/modeling/losses/probiou_loss.py

+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+import paddle
+import paddle.nn.functional as F
+
+from ppdet.core.workspace import register, serializable
+
+__all__ = ['ProbIoULoss']
+
+
+def gbb_form(boxes):
+    xy, wh, angle = paddle.split(boxes, [2, 2, 1], axis=-1)
+    return paddle.concat([xy, wh.pow(2) / 12., angle], axis=-1)
+
+
+def rotated_form(a_, b_, angles):
+    cos_a = paddle.cos(angles)
+    sin_a = paddle.sin(angles)
+    a = a_ * paddle.pow(cos_a, 2) + b_ * paddle.pow(sin_a, 2)
+    b = a_ * paddle.pow(sin_a, 2) + b_ * paddle.pow(cos_a, 2)
+    c = (a_ - b_) * cos_a * sin_a
+    return a, b, c
+
+
+def probiou_loss(pred, target, eps=1e-3, mode='l1'):
+    """
+        pred    -> a matrix [N,5](x,y,w,h,angle - in radians) containing ours predicted box ;in case of HBB angle == 0
+        target  -> a matrix [N,5](x,y,w,h,angle - in radians) containing ours target    box ;in case of HBB angle == 0
+        eps     -> threshold to avoid infinite values
+        mode    -> ('l1' in [0,1] or 'l2' in [0,inf]) metrics according our paper
+
+    """
+
+    gbboxes1 = gbb_form(pred)
+    gbboxes2 = gbb_form(target)
+
+    x1, y1, a1_, b1_, c1_ = gbboxes1[:,
+                                     0], gbboxes1[:,
+                                                  1], gbboxes1[:,
+                                                               2], gbboxes1[:,
+                                                                            3], gbboxes1[:,
+                                                                                         4]
+    x2, y2, a2_, b2_, c2_ = gbboxes2[:,
+                                     0], gbboxes2[:,
+                                                  1], gbboxes2[:,
+                                                               2], gbboxes2[:,
+                                                                            3], gbboxes2[:,
+                                                                                         4]
+
+    a1, b1, c1 = rotated_form(a1_, b1_, c1_)
+    a2, b2, c2 = rotated_form(a2_, b2_, c2_)
+
+    t1 = 0.25 * ((a1 + a2) * (paddle.pow(y1 - y2, 2)) + (b1 + b2) * (paddle.pow(x1 - x2, 2))) + \
+         0.5 * ((c1+c2)*(x2-x1)*(y1-y2))
+    t2 = (a1 + a2) * (b1 + b2) - paddle.pow(c1 + c2, 2)
+    t3_ = (a1 * b1 - c1 * c1) * (a2 * b2 - c2 * c2)
+    t3 = 0.5 * paddle.log(t2 / (4 * paddle.sqrt(F.relu(t3_)) + eps))
+
+    B_d = (t1 / t2) + t3
+    # B_d = t1 + t2 + t3
+
+    B_d = paddle.clip(B_d, min=eps, max=100.0)
+    l1 = paddle.sqrt(1.0 - paddle.exp(-B_d) + eps)
+    l_i = paddle.pow(l1, 2.0)
+    l2 = -paddle.log(1.0 - l_i + eps)
+
+    if mode == 'l1':
+        probiou = l1
+    if mode == 'l2':
+        probiou = l2
+
+    return probiou
+
+
+@serializable
+@register
+class ProbIoULoss(object):
+    """ ProbIoU Loss, refer to https://arxiv.org/abs/2106.06072 for details """
+
+    def __init__(self, mode='l1', eps=1e-3):
+        super(ProbIoULoss, self).__init__()
+        self.mode = mode
+        self.eps = eps
+
+    def __call__(self, pred_rboxes, assigned_rboxes):
+        return probiou_loss(pred_rboxes, assigned_rboxes, self.eps, self.mode)

ppdet/modeling/ops.py

@@ -37,6 +37,7 @@ __all__ = [
     'silu',
     'swish',
     'identity',
+    'anchor_generator'
 ]
 
 
@@ -117,6 +118,101 @@ def batch_norm(ch,
 
     return norm_layer
 
+@paddle.jit.not_to_static
+def anchor_generator(input,
+                     anchor_sizes=None,
+                     aspect_ratios=None,
+                     variance=[0.1, 0.1, 0.2, 0.2],
+                     stride=None,
+                     offset=0.5):
+    """
+    **Anchor generator operator**
+    Generate anchors for Faster RCNN algorithm.
+    Each position of the input produce N anchors, N =
+    size(anchor_sizes) * size(aspect_ratios). The order of generated anchors
+    is firstly aspect_ratios loop then anchor_sizes loop.
+    Args:
+       input(Variable): 4-D Tensor with shape [N,C,H,W]. The input feature map.
+       anchor_sizes(float32|list|tuple, optional): The anchor sizes of generated
+          anchors, given in absolute pixels e.g. [64., 128., 256., 512.].
+          For instance, the anchor size of 64 means the area of this anchor 
+          equals to 64**2. None by default.
+       aspect_ratios(float32|list|tuple, optional): The height / width ratios 
+           of generated anchors, e.g. [0.5, 1.0, 2.0]. None by default.
+       variance(list|tuple, optional): The variances to be used in box 
+           regression deltas. The data type is float32, [0.1, 0.1, 0.2, 0.2] by 
+           default.
+       stride(list|tuple, optional): The anchors stride across width and height.
+           The data type is float32. e.g. [16.0, 16.0]. None by default.
+       offset(float32, optional): Prior boxes center offset. 0.5 by default.
+    Returns:
+        Tuple:
+        Anchors(Variable): The output anchors with a layout of [H, W, num_anchors, 4].
+        H is the height of input, W is the width of input,
+        num_anchors is the box count of each position. 
+        Each anchor is in (xmin, ymin, xmax, ymax) format an unnormalized.
+ 
+        Variances(Variable): The expanded variances of anchors
+        with a layout of [H, W, num_priors, 4].
+        H is the height of input, W is the width of input
+        num_anchors is the box count of each position.
+        Each variance is in (xcenter, ycenter, w, h) format.
+    Examples:
+        .. code-block:: python
+            import paddle.fluid as fluid
+            conv1 = fluid.data(name='conv1', shape=[None, 48, 16, 16], dtype='float32')
+            anchor, var = fluid.layers.anchor_generator(
+                input=conv1,
+                anchor_sizes=[64, 128, 256, 512],
+                aspect_ratios=[0.5, 1.0, 2.0],
+                variance=[0.1, 0.1, 0.2, 0.2],
+                stride=[16.0, 16.0],
+                offset=0.5)
+    """
+
+    def _is_list_or_tuple_(data):
+        return (isinstance(data, list) or isinstance(data, tuple))
+
+    if not _is_list_or_tuple_(anchor_sizes):
+        anchor_sizes = [anchor_sizes]
+    if not _is_list_or_tuple_(aspect_ratios):
+        aspect_ratios = [aspect_ratios]
+    if not (_is_list_or_tuple_(stride) and len(stride) == 2):
+        raise ValueError('stride should be a list or tuple ',
+                         'with length 2, (stride_width, stride_height).')
+
+    anchor_sizes = list(map(float, anchor_sizes))
+    aspect_ratios = list(map(float, aspect_ratios))
+    stride = list(map(float, stride))
+
+    if in_dynamic_mode():
+        attrs = ('anchor_sizes', anchor_sizes, 'aspect_ratios', aspect_ratios,
+                 'variances', variance, 'stride', stride, 'offset', offset)
+        anchor, var = C_ops.anchor_generator(input, *attrs)
+        return anchor, var
+
+    helper = LayerHelper("anchor_generator", **locals())
+    dtype = helper.input_dtype()
+    attrs = {
+        'anchor_sizes': anchor_sizes,
+        'aspect_ratios': aspect_ratios,
+        'variances': variance,
+        'stride': stride,
+        'offset': offset
+    }
+
+    anchor = helper.create_variable_for_type_inference(dtype)
+    var = helper.create_variable_for_type_inference(dtype)
+    helper.append_op(
+        type="anchor_generator",
+        inputs={"Input": input},
+        outputs={"Anchors": anchor,
+                 "Variances": var},
+        attrs=attrs, )
+    anchor.stop_gradient = True
+    var.stop_gradient = True
+    return anchor, var
+
 
 @paddle.jit.not_to_static
 def distribute_fpn_proposals(fpn_rois,

ppdet/modeling/rbox_utils.py

@@ -157,3 +157,85 @@ def rbox2poly_np(rboxes):
         polys.append(poly)
     polys = np.array(polys)
     return polys
+
+
+# rbox function implemented using paddle
+def box2corners(box):
+    """convert box coordinate to corners
+    Args:
+        box (Tensor): (B, N, 5) with (x, y, w, h, alpha) angle is in [0, 90)
+    Returns:
+        corners (Tensor): (B, N, 4, 2) with (x1, y1, x2, y2, x3, y3, x4, y4)
+    """
+    B = box.shape[0]
+    x, y, w, h, alpha = paddle.split(box, 5, axis=-1)
+    x4 = paddle.to_tensor(
+        [0.5, 0.5, -0.5, -0.5], dtype=paddle.float32).reshape(
+            (1, 1, 4))  # (1,1,4)
+    x4 = x4 * w  # (B, N, 4)
+    y4 = paddle.to_tensor(
+        [-0.5, 0.5, 0.5, -0.5], dtype=paddle.float32).reshape((1, 1, 4))
+    y4 = y4 * h  # (B, N, 4)
+    corners = paddle.stack([x4, y4], axis=-1)  # (B, N, 4, 2)
+    sin = paddle.sin(alpha)
+    cos = paddle.cos(alpha)
+    row1 = paddle.concat([cos, sin], axis=-1)
+    row2 = paddle.concat([-sin, cos], axis=-1)  # (B, N, 2)
+    rot_T = paddle.stack([row1, row2], axis=-2)  # (B, N, 2, 2)
+    rotated = paddle.bmm(corners.reshape([-1, 4, 2]), rot_T.reshape([-1, 2, 2]))
+    rotated = rotated.reshape([B, -1, 4, 2])  # (B*N, 4, 2) -> (B, N, 4, 2)
+    rotated[..., 0] += x
+    rotated[..., 1] += y
+    return rotated
+
+
+def paddle_gather(x, dim, index):
+    index_shape = index.shape
+    index_flatten = index.flatten()
+    if dim < 0:
+        dim = len(x.shape) + dim
+    nd_index = []
+    for k in range(len(x.shape)):
+        if k == dim:
+            nd_index.append(index_flatten)
+        else:
+            reshape_shape = [1] * len(x.shape)
+            reshape_shape[k] = x.shape[k]
+            x_arange = paddle.arange(x.shape[k], dtype=index.dtype)
+            x_arange = x_arange.reshape(reshape_shape)
+            dim_index = paddle.expand(x_arange, index_shape).flatten()
+            nd_index.append(dim_index)
+    ind2 = paddle.transpose(paddle.stack(nd_index), [1, 0]).astype("int64")
+    paddle_out = paddle.gather_nd(x, ind2).reshape(index_shape)
+    return paddle_out
+
+
+def check_points_in_polys(points, polys):
+    """Check whether point is in rotated boxes
+    Args:
+        points (tensor): (1, L, 2) anchor points
+        polys (tensor): [B, N, 4, 2] gt_polys
+        eps (float): default 1e-9
+    Returns:
+        is_in_polys (tensor): (B, N, L)
+    """
+    # [1, L, 2] -> [1, 1, L, 2]
+    points = points.unsqueeze(0)
+    # [B, N, 4, 2] -> [B, N, 1, 2]
+    a, b, c, d = polys.split(4, axis=2)
+    ab = b - a
+    ad = d - a
+    # [B, N, L, 2]
+    ap = points - a
+    # [B, N, 1]
+    norm_ab = paddle.sum(ab * ab, axis=-1)
+    # [B, N, 1]
+    norm_ad = paddle.sum(ad * ad, axis=-1)
+    # [B, N, L] dot product
+    ap_dot_ab = paddle.sum(ap * ab, axis=-1)
+    # [B, N, L] dot product
+    ap_dot_ad = paddle.sum(ap * ad, axis=-1)
+    # [B, N, L] <A, B> = |A|*|B|*cos(theta)
+    is_in_polys = (ap_dot_ab >= 0) & (ap_dot_ab <= norm_ab) & (
+        ap_dot_ad >= 0) & (ap_dot_ad <= norm_ad)
+    return is_in_polys