- We use a multi-scale evaluation strategy to get the mAP in `Easy/Medium/Hard Set`. Please refer to the [evaluation on the WIDER FACE dataset](#Evaluated-on-the-WIDER-FACE-Dataset) for details.
We provide standalone deploy of PaddleInference(Server-GPU)、PaddleLite(mobile、ARM)、Third-Engine(MNN、OpenVino), which is independent of training codes。For detail, please click [Deploy-docs](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/deploy/README_en.md)。
We provide standalone deploy of PaddleInference(Server-GPU)、PaddleLite(mobile、ARM)、Third-Engine(MNN、OpenVino), which is independent of training codes。For detail, please click [Deploy-docs](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/deploy/README_en.md)。
## Train with custom data
...
...
@@ -225,7 +225,7 @@ For more configs, please refer to [KeyPointConfigGuide](../../docs/tutorials/Key
## BenchMark
We provide benchmarks in different runtime environments for your reference when choosing models. See [Keypoint Inference Benchmark](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/keypoint/KeypointBenchmark.md) for details.
We provide benchmarks in different runtime environments for your reference when choosing models. See [Keypoint Inference Benchmark](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/configs/keypoint/KeypointBenchmark.md) for details.
-<aname="latency">Latency:</a> All our models test on `Intel core i7 10750H` CPU with MKLDNN by 12 threads and `Qualcomm Snapdragon 865(4xA77+4xA55)` with 4 threads by arm8 and with FP16. In the above table, test CPU latency on Paddle-Inference and testing Mobile latency with `Lite`->[Paddle-Lite](https://github.com/PaddlePaddle/Paddle-Lite).
- PicoDet is trained on COCO train2017 dataset and evaluated on COCO val2017. And PicoDet used 4 GPUs for training and all checkpoints are trained with default settings and hyperparameters.
- Benchmark test: When testing the speed benchmark, the post-processing is not included in the exported model, you need to set `-o export.benchmark=True` or manually modify [runtime.yml](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/runtime.yml#L12).
- Benchmark test: When testing the speed benchmark, the post-processing is not included in the exported model, you need to set `-o export.benchmark=True` or manually modify [runtime.yml](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/configs/runtime.yml#L12).
</details>
...
...
@@ -82,8 +82,8 @@ We developed a series of lightweight models, named `PP-PicoDet`. Because of the
- If no post processing is required, please specify: `-o export.benchmark=True` (if -o has already appeared, delete -o here) or manually modify corresponding fields in [runtime.yml](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/runtime.yml).
- If no NMS is required, please specify: `-o export.nms=True` or manually modify corresponding fields in [runtime.yml](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/runtime.yml). Many scenes exported to ONNX only support single input and fixed shape output, so if exporting to ONNX, it is recommended not to export NMS.
- If no post processing is required, please specify: `-o export.benchmark=True` (if -o has already appeared, delete -o here) or manually modify corresponding fields in [runtime.yml](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/configs/runtime.yml).
- If no NMS is required, please specify: `-o export.nms=True` or manually modify corresponding fields in [runtime.yml](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/configs/runtime.yml). Many scenes exported to ONNX only support single input and fixed shape output, so if exporting to ONNX, it is recommended not to export NMS.
Please refer this [documentation](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/picodet/legacy_model/pruner/README.md) for details such as requirements, training and deployment.
Please refer this [documentation](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.6/configs/picodet/legacy_model/pruner/README.md) for details such as requirements, training and deployment.
</details>
## Application
-**Pedestrian detection:** model zoo of `PicoDet-S-Pedestrian` please refer to [PP-TinyPose](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/keypoint/tiny_pose#%E8%A1%8C%E4%BA%BA%E6%A3%80%E6%B5%8B%E6%A8%A1%E5%9E%8B)
-**Pedestrian detection:** model zoo of `PicoDet-S-Pedestrian` please refer to [PP-TinyPose](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.6/configs/keypoint/tiny_pose#%E8%A1%8C%E4%BA%BA%E6%A3%80%E6%B5%8B%E6%A8%A1%E5%9E%8B)
-**Mainbody detection:** model zoo of `PicoDet-L-Mainbody` please refer to [mainbody detection](./legacy_model/application/mainbody_detection/README.md)
...
...
@@ -315,7 +315,7 @@ Please use `PicoDet-LCNet` model, which has fewer `transpose` operators.
<details>
<summary>How to count model parameters.</summary>
You can insert below code at [here](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/ppdet/engine/trainer.py#L141) to count learnable parameters.
You can insert below code at [here](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/ppdet/engine/trainer.py#L141) to count learnable parameters.
@@ -17,7 +17,7 @@ The network for detecting vehicles is YOLOv3, the backbone of which is Dacknet53
### 2. Configuration for training
PaddleDetection provides users with a configuration file [yolov3_darknet53_270e_coco.yml](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) to train YOLOv3 on the COCO dataset, compared with this file, we modify some parameters as followed to conduct the training for pedestrian detection:
PaddleDetection provides users with a configuration file [yolov3_darknet53_270e_coco.yml](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/configs/yolov3/yolov3_darknet53_270e_coco.yml) to train YOLOv3 on the COCO dataset, compared with this file, we modify some parameters as followed to conduct the training for pedestrian detection:
@@ -17,7 +17,7 @@ The network for detecting vehicles is YOLOv3, the backbone of which is Dacknet53
### 2. Configuration for training
PaddleDetection provides users with a configuration file [yolov3_darknet53_270e_coco.yml](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) to train YOLOv3 on the COCO dataset, compared with this file, we modify some parameters as followed to conduct the training for vehicle detection:
PaddleDetection provides users with a configuration file [yolov3_darknet53_270e_coco.yml](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/configs/yolov3/yolov3_darknet53_270e_coco.yml) to train YOLOv3 on the COCO dataset, compared with this file, we modify some parameters as followed to conduct the training for vehicle detection:
| Backbone | Network type | Number of images per GPU | Learning rate strategy | Inferring time(fps) | Box AP | Mask AP | Download | Configuration File |
- 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>)**.
**Attention:** Due to the overall upgrade of the dynamic graph framework, the weighting model published by PaddleDetection of PAF Net needs to be evaluated with a --bias field, for example
@@ -143,13 +143,13 @@ For rtsp pull stream, use --rtsp RTSP [RTSP ...] parameter to specify one or mor
- rtsp push stream
For rtsp push stream, use --pushurl rtsp:[IP] parameter to push stream to a IP set, and you can visualize the output video by [VLC Player](https://vlc.onl/) with the `open network` funciton. the whole url path is `rtsp:[IP]/videoname`, the videoname here is the basename of the video file to infer, and the default of videoname is `output` when the video come from local camera and there is no video name.
For rtsp push stream, use --pushurl rtsp:[IP] parameter to push stream to a IP set, and you can visualize the output video by [VLC Player](https://vlc.onl/) with the `open network` funciton. the whole url path is `rtsp:[IP]/videoname`, the videoname here is the basename of the video file to infer, and the default of videoname is `output` when the video come from local camera and there is no video name.
```
# Example：license plate recognition，in this example the whole url path is: rtsp://[YOUR_SERVER_IP]:8554/test_video
1. rtsp push stream is based on [rtsp-simple-server](https://github.com/aler9/rtsp-simple-server), please enable this serving first.
2. the output visualize will be frozen frequently if the model cost too much time, we suggest to use faster model like ppyoloe_s in tracking, this is simply replace mot_ppyoloe_l_36e_pipeline.zip with mot_ppyoloe_s_36e_pipeline.zip in model config yaml file.
- [Fixed Virtual Platform (FVP) based on Arm(R) Corstone(TM)-300 software](https://developer.arm.com/tools-and-software/open-source-software/arm-platforms-software/arm-ecosystem-fvps)
- [GCC toolchain from Arm(R)](https://developer.arm.com/-/media/Files/downloads/gnu-rm/10-2020q4/gcc-arm-none-eabi-10-2020-q4-major-x86_64-linux.tar.bz2)
- Release rotated object detector PP-YOLOE-R：SOTA Anchor-free rotated object detection model with high accuracy and efficiency. It has a series of models, named s/m/l/x, for cloud and edge devices and avoids using special operators to be deployed friendly with TensorRT.
- Release small object detector PP-YOLOE-SOD: End-to-end detection pipeline based on sliced images and SOTA model on VisDrone based on original images.
- Release crowded object detector: Crowded object detection model with top accuracy on SKU dataset.
- Functions in different scenarios
- Release real-time object detection model on edge device in PP-Human v2. The model reaches 45.7mAP and 80FPS on Jetson AGX
- Release real-time object detection model on edge device in PP-Vehicle. The model reaches 53.5mAP and 80FPS on Jetson AGX
- Support multi-stream deployment in PP-Human v2 and PP-Vehicle. Achieved 20FPS in 4-stream deployment on Jetson AGX
- Support retrograde and press line detection in PP-Vehicle
- Cutting-edge algorithms
- Release YOLOv8 and YOLOv6 3.0 in YOLO Family
- Release object detection algorithm DINO, YOLOF
- Rich ViTDet series including PP-YOLOE+ViT_base, Mask RCNN + ViT_base, Mask RCNN + ViT_large
@@ -170,7 +170,7 @@ At this point, this model can be used in PP-Human.
In the model of action recognition based on detection with human id, the task is defined to detect target objects in images of corresponding person. When the target object is detected, the behavior type of the character in a certain period of time. The type of the corresponding classification is regarded as the action of the current period. Therefore, on the basis of completing the training and deployment of the custom model, it is also necessary to convert the detection model results to the final action recognition results as output, and the displayed result of the visualization should be modified.
#### Convert to Action Recognition Result
Please modify the [postprocessing function](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/deploy/pipeline/pphuman/action_infer.py#L338).
Please modify the [postprocessing function](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/deploy/pipeline/pphuman/action_infer.py#L338).
The core code are:
```python
...
...
@@ -196,4 +196,4 @@ else:
```
#### Modify Visual Output
At present, ID-based action recognition is displayed based on the results of action recognition and predefined category names. For the detail, please refer to [here](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/deploy/pipeline/pipeline.py#L1024-L1043). If the custom action needs to be modified to another display name, please modify it accordingly to output the corresponding result.
At present, ID-based action recognition is displayed based on the results of action recognition and predefined category names. For the detail, please refer to [here](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/deploy/pipeline/pipeline.py#L1024-L1043). If the custom action needs to be modified to another display name, please modify it accordingly to output the corresponding result.
In the skeleton-based action recognition, the classification result of the model represents the behavior type of the character in a certain period of time. The type of the corresponding classification is regarded as the action of the current period. Therefore, on the basis of completing the training and deployment of the custom model, the model output is directly used as the final result, and the displayed result of the visualization should be modified.
#### Modify Visual Output
At present, ID-based action recognition is displayed based on the results of action recognition and predefined category names. For the detail, please refer to [here](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/deploy/pipeline/pipeline.py#L1024-L1043). If the custom action needs to be modified to another display name, please modify it accordingly to output the corresponding result.
At present, ID-based action recognition is displayed based on the results of action recognition and predefined category names. For the detail, please refer to [here](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/deploy/pipeline/pipeline.py#L1024-L1043). If the custom action needs to be modified to another display name, please modify it accordingly to output the corresponding result.
@@ -60,7 +60,7 @@ Augmentation of covered data in keypoint model training to improve model perform
The keypoint model is trained and predicted on the basis of image, and video input is also predicted by splitting the video into frames. Although the content is mostly similar between frames, small differences may still lead to large changes in the output of the model. As a result of that, although the predicted coordinates are roughly correct, there may be jitters in the visual effect.
By adding a smoothing filter process, the performance of the video output can be effectively improved by combining the predicted results of each frame and the historical results. For this part, please see [Filter Smoothing](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/deploy/python/det_keypoint_unite_infer.py#L206).
By adding a smoothing filter process, the performance of the video output can be effectively improved by combining the predicted results of each frame and the historical results. For this part, please see [Filter Smoothing](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/deploy/python/det_keypoint_unite_infer.py#L206).
@@ -223,7 +223,7 @@ The same applies to the deletion of attributes.
After modifying the attribute definition, the post-processing part of the pipeline also needs to be modified accordingly, which mainly affects the display results when the results are visualized.
The code is at [file](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/deploy/pipeline/ppvehicle/vehicle_attr.py#L108), that is, the `postprocess` function.
The code is at [file](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/deploy/pipeline/ppvehicle/vehicle_attr.py#L108), that is, the `postprocess` function.
The function implementation is described as follows:
- For more CUDA version or environment to quick install, please refer to the [PaddlePaddle Quick Installation document](https://www.paddlepaddle.org.cn/install/quick)
@@ -147,7 +147,7 @@ Then, it will generate following contents:
#### Annotation file(json)—>COCO Dataset
Convert the data annotated by LabelMe to COCO dataset by the script [x2coco.py](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/tools/x2coco.py) provided by PaddleDetection.
Convert the data annotated by LabelMe to COCO dataset by the script [x2coco.py](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/tools/x2coco.py) provided by PaddleDetection.
**PaddleDetection supports the format of VOC or COCO.** The annotation file generated by LabelImg needs to be converted by VOC or COCO. You can refer to [PrepareDataSet](./PrepareDataSet.md#%E5%87%86%E5%A4%87%E8%AE%AD%E7%BB%83%E6%95%B0%E6%8D%AE).
Convert the data annotated by LabelMe to COCO dataset by this script [x2coco.py](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/tools/x2coco.py).
Convert the data annotated by LabelMe to COCO dataset by this script [x2coco.py](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/tools/x2coco.py).
