update vot code (#4338)

PaddleCV/tracking/README.md

 # tracking 单目标跟踪框架
 
-## 目标跟踪介绍
+## 介绍
 
 tracking 是基于百度深度学习框架Paddle研发的视频单目标跟踪（Visual Object Tracking, VOT）库, 整体框架参考 [pytracking](https://github.com/visionml/pytracking)，其优秀的设计使得我们能够方便地将其他跟踪器如SiamFC，SiamRPN，SiamMask等融合到一个框架中，方便后续统一的实验和比较。
 
 当前tracking涵盖当前目标跟踪的主流模型，包括SiamFC, SiamRPN, SiamMask, ATOM。tracking旨在给开发者提供一系列基于PaddlePaddle的便捷、高效的目标跟踪深度学习算法，后续会不断的扩展模型的丰富度。
 
-## 目标跟踪库的代码目录结构
+## 代码目录结构
 
 
 ```
@@ -36,7 +36,7 @@ pytracking  包含跟踪代码
 
 ### 数据准备
 
-目标跟踪的训练集和测试集是不同的，目前最好的模型往往是使用多个训练集进行训练。常用的数据集如下:
+目标跟踪的训练集和测试集是不同的，目前最好的模型往往是使用多个训练集进行训练。
 
 主流的训练数据集有：
 - [VID](http://bvisionweb1.cs.unc.edu/ilsvrc2015/ILSVRC2015_VID.tar.gz)
@@ -60,6 +60,8 @@ tracking的工作环境：
 - python3
 - PaddlePaddle1.7
 
+> 注意：如果遇到cmath无法import的问题，建议切换Python版本，建议使用python3.6.8, python3.7.0
+
 ### 安装依赖
 
 1. 安装paddle，需要安装1.7版本的Paddle，如低于这个版本，请升级到Paddle 1.7.

PaddleCV/tracking/pytracking/eval_benchmark.py

@@ -19,9 +19,9 @@ CURRENT_DIR = osp.dirname(__file__)
 sys.path.append(osp.join(CURRENT_DIR, '..'))
 
 from pytracking.admin.environment import env_settings
-from pytracking.pysot_toolkit.datasets import DatasetFactory
-from pytracking.pysot_toolkit.evaluation import EAOBenchmark, AccuracyRobustnessBenchmark, OPEBenchmark
-from pytracking.pysot_toolkit.utils.region import vot_overlap
+from pytracking.pysot_toolkit.pysot.datasets import DatasetFactory
+from pytracking.pysot_toolkit.pysot.evaluation import EAOBenchmark, AccuracyRobustnessBenchmark, OPEBenchmark
+from pytracking.pysot_toolkit.pysot.utils.region import vot_overlap
 
 parser = argparse.ArgumentParser(description='tracking evaluation')
 

PaddleCV/tracking/pytracking/libs/optimization.py

@@ -394,7 +394,7 @@ class ConjugateGradient(ConjugateGradientBase):
                                fetch_list=[v.name for v in self.f0])
             res = TensorList(res)
             loss = self.problem.ip_output(res, res)
-            print('Paddle Loss: {}'.format(loss))
+            #print('Paddle Loss: {}'.format(loss))
 
 
 class GaussNewtonCG(ConjugateGradientBase):
@@ -614,7 +614,7 @@ class GaussNewtonCG(ConjugateGradientBase):
                                fetch_list=[v.name for v in self.f0])
             res = TensorList(res)
             loss = self.problem.ip_output(res, res)
-            print('Paddle Loss: {}'.format(loss))
+            #print('Paddle Loss: {}'.format(loss))
 
 
 class GradientDescentL2:
@@ -691,7 +691,7 @@ class GradientDescentL2:
                                fetch_list=[self.loss.name] + grad_names)
             if self.debug:
                 loss = res[0]
-                print('Paddle Loss: {}'.format(loss))
+                #print('Paddle Loss: {}'.format(loss))
 
             grad = TensorList(res[1:])
 

PaddleCV/tracking/pytracking/parameter/atom/default_vot.py

 import numpy as np
 
-from pytracking.features.deep import ResNet50
+from pytracking.features.deep import ResNet18, ResNet50
 from pytracking.features.extractor import MultiResolutionExtractor
 from pytracking.utils import TrackerParams, FeatureParams
 
@@ -42,7 +42,9 @@ def parameters():
     params.train_skipping = 10  # How often to run training (every n-th frame)
 
     # Online model parameters
-    deep_params.kernel_size = (4, 4)  # Kernel size of filter
+    # deep_params.kernel_size = (4, 4) # when slice double grad is support,
+    # else, deep_params.kernel_size = (5, 5)
+    deep_params.kernel_size = (5, 5)  # Kernel size of filter
     deep_params.compressed_dim = 64  # Dimension output of projection matrix
     deep_params.filter_reg = 1e-1  # Filter regularization factor
     deep_params.projection_reg = 1e-4  # Projection regularization factor
@@ -104,11 +106,8 @@ def parameters():
 
     # Setup the feature extractor (which includes the IoUNet)
     deep_fparams = FeatureParams(feature_params=[deep_params])
-    deep_feat = ResNet50(
-        net_path='/home/vis/bily/code/baidu/personal-code/libi-13/paddle_ATOMnet-ep0040',
-        output_layers=['block2'],
-        fparams=deep_fparams,
-        normalize_power=2)
+    deep_feat = ResNet18(
+        output_layers=['block2'], fparams=deep_fparams, normalize_power=2)
     params.features = MultiResolutionExtractor([deep_feat])
 
     params.vot_anno_conversion_type = 'preserve_area'