update tracking

README.md

@@ -65,7 +65,7 @@ More results and models are available in the [docs/MODEL_ZOO.md](docs/MODEL_ZOO.
     <img src="docs/posetrack2.gif", width="344">
 </p>
 
-Please read [PoseFlow/README.md](PoseFlow/) for details.
+Please read [trackers/README.md](trackers/) for details.
 
 ### CrowdPose
 <p align='center'>

alphapose/utils/vis.py

@@ -136,7 +136,7 @@ def vis_frame_fast(frame, im_res, opt, format='coco'):
             if 'box' in human.keys():
                 bbox = human['box']
             else:
-                from PoseFlow.poseflow_infer import get_box
+                from trackers.PoseFlow.poseflow_infer import get_box
                 keypoints = []
                 for n in range(kp_scores.shape[0]):
                     keypoints.append(float(kp_preds[n, 0]))
@@ -284,7 +284,7 @@ def vis_frame(frame, im_res, opt, format='coco'):
                 bbox = human['box']
                 bbox = [bbox[0], bbox[0]+bbox[2], bbox[1], bbox[1]+bbox[3]]#xmin,xmax,ymin,ymax
             else:
-                from PoseFlow.poseflow_infer import get_box
+                from trackers.PoseFlow.poseflow_infer import get_box
                 keypoints = []
                 for n in range(kp_scores.shape[0]):
                     keypoints.append(float(kp_preds[n, 0]))

alphapose/utils/writer.py

@@ -44,7 +44,7 @@ def __init__(self, cfg, opt, save_video=False,
                 os.mkdir(opt.outputpath + '/vis')
 
         if opt.pose_flow:
-            from PoseFlow.poseflow_infer import PoseFlowWrapper
+            from trackers.PoseFlow.poseflow_infer import PoseFlowWrapper
             self.pose_flow_wrapper = PoseFlowWrapper(save_path=os.path.join(opt.outputpath, 'poseflow'))
 
     def start_worker(self, target):

docs/run.md

@@ -25,11 +25,14 @@ Here, we first list the flags and other parameters you can tune. Default paramet
 - `--min_box_area`: Min box area to filter out, you can set it like 100 to filter out small people.
 - `--gpus`: Choose which cuda device to use by index and input comma to use multi gpus, e.g. 0,1,2,3. (input -1 for cpu only)
 
+- `--pose_track`: Enable tracking pipeline with human re-id feature, it is currently the best performance pose tracker
+- `--pose_flow`: This flag will be depreciated. It enables the old tracking version of PoseFlow.
+
 All the flags available here: [link](../scripts/demo_inference.py#L22)
 
 
 ## Parameters
 1. yolo detector config is [here](../detector/yolo_cfg.py)
 - `CONFIDENCE`: Confidence threshold for human detection. Lower the value can improve the final accuracy but decrease the speed. Default is 0.05.
 - `NMS_THRES`: NMS threshold for human detection. Increase the value can improve the final accuracy but decrease the speed. Default is 0.6.
-- `INP_DIM`: The input size of detection network. The inp_dim should be multiple of 32. Default is 608. Increase it may improve the accuracy.
+- `INP_DIM`: The input size of detection network. The inp_dim should be multiple of 32. Default is 608. Increase it may improve the accuracy.

PoseFlow/tracker-general.py → trackers/PoseFlow/tracker-general.py

@@ -175,7 +175,7 @@ def load_pose_boxes(img_name):
             tasks.append((img1_path,img2_path, image_dir, frame_id, next_frame_id))
 
     # do the matching parallel
-    parallel_process(tasks, orb_matching, n_jobs=16)
+    parallel_process(tasks, orb_matching, n_jobs=8)
 
     print("Start pose tracking...\n")
     # tracking process

trackers/README.md

@@ -1,18 +1,34 @@
-# Pose Tracking
-## Models
+# Pose Tracking Module for AlphaPose
+
+
+## Human-ReID based tracking (Recommended)
+Currently the best performance tracking model. Paper coming soon.
+
+### Getting started
 Download  [human reid model](https://mega.nz/#!YTZFnSJY!wlbo_5oa2TpDAGyWCTKTX1hh4d6DvJhh_RUA2z6i_so) and place it into `./trackers/weights/`.
 
+Then simply run alphapose with additional flag `--pose_track`
+
 You can try different person reid model by modifing `cfg.arch` and `cfg.loadmodel` in `./trackers/tracker_cfg.py`.
 
 If you want to train your own reid model, please refer to this [project](https://github.com/KaiyangZhou/deep-person-reid)
-## Demo
+
+### Demo
 ``` bash
 ./scripts/inference.sh ${CONFIG} ${CHECKPOINT} ${VIDEO_NAME}  ${OUTPUT_DIR}, --pose_track
 ```
-## Todo
+### Todo
 - [] Evaluation Tools for PoseTrack
 - [] More Models
 - [] Training code for [PoseTrack Dataset](https://posetrack.net/)
 
 
+## PoseFlow human tracking
+This tracker is based on our BMVC 2018 paper PoseFlow.
+
+### Getting started
+
+Simply run alphapose with additional flag `--pose_flow`
 
+### More info
+For more info, please refer to [PoseFlow/README.md](PoseFlow/)