Awesome

Tracking People by Predicting 3D Appearance, Location & Pose

Code repository for the paper "Tracking People by Predicting 3D Appearance, Location & Pose".
Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, Jitendra Malik.

This code repository provides a code implementation for our paper PHALP, with installation, a demo code to run on any videos, preparing datasets, and evaluating on datasets.

<b>This branch contains code supporting our latest work: 4D-Humans. <br> For the original PHALP code, please see the initial release branch.</b>

Installation

After installing the PyTorch dependency, you may install our phalp package directly as:

pip install phalp[all]@git+https://github.com/brjathu/PHALP.git

git clone https://github.com/brjathu/PHALP.git
cd PHALP
conda create -n phalp python=3.10
conda activate phalp
conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia
pip install -e .[all]

Demo

To run our code on a video, please specifiy the input video video.source and an output directory video.output_dir:

python scripts/demo.py video.source=assets/videos/gymnasts.mp4 video.output_dir='outputs'

The output directory will contain a video rendering of the tracklets and a .pkl file containing the tracklets with 3D pose and shape (see structure below). <br>

Command-line options

Input Sources

You can specify various kinds of input sources. For example, you can specify a video file, a youtube video, a directory of images:

# for a video file
python scripts/demo.py video.source=assets/videos/vid.mp4

# for a youtube video
python scripts/demo.py video.source=\'"https://www.youtube.com/watch?v=xEH_5T9jMVU"\'

# for a directory of images
python scripts/demo.py video.source=<dirtory_path>

In addition to these options, you can also give images and bounding boxes as inputs, so the model will only do tracking using the given bounding boxes. To do this, you need to specify the video.source as a .pkl file, where each key is the frame name and the absolute path to the image is computed as os.path.join(video.base_path, frame_name). The value of each key is a dictionary with the following keys: gt_bbox, gt_class, gt_track_id. Please see the following example. gt_boxes is a np.ndarray of shape (N, 4) where each row is a bounding box in the format of [x1, y1, x2, y2]. You can also give gt_class and gt_track_id to store it in the final output.

 gt_data[frame_id] = {
                        "gt_bbox": gt_boxes,
                        "extra_data": {
                            "gt_class": [],
                            "gt_track_id": [],
                        }
                    }

Here is an example, of how to give bounding boxes and track-ids to the model and get the renderings.

mkdir assets/videos/gymnasts
ffmpeg -i assets/videos/gymnasts.mp4 -q:v 2 assets/videos/gymnasts/%06d.jpg

python scripts/demo.py \
render.enable=True \
video.output_dir=test_gt_bbox \
use_gt=True \
video.base_path=assets/videos/gymnasts \
video.source=assets/videos/gt_tracks.pkl

Running on a subset of frames

You can specify the start and end of the video to be tracked, e.g. track from frame 50 to 100:

python scripts/demo.py video.source=assets/videos/vid.mp4 video.start_frame=50 video.end_frame=100

However, if the video is too long and extracting the frames is too time consuming, you can set video.extract_video=False. This will use the torchvision backend and it will only keep the timestamps of the video in memeory. If this is enabled, you can give start time and end time of the video in seconds.

python scripts/demo.py video.source=assets/videos/vid.mp4 video.extract_video=False video.start_time=1s video.end_time=2s

Visualization type

We support multiple types of visualization in render.type: HUMAN_MESH (default) renders the full human mesh, HUMAN_MASK visualizes the segmentation masks, HUMAN_BBOX visualizes the bounding boxes with track-ids, TRACKID_<id>_MESH renders the full human mesh but for track <id> only:

# render full human mesh
python scripts/demo.py video.source=assets/videos/vid.mp4 render.type=HUMAN_MESH

# render segmentation mask
python scripts/demo.py video.source=assets/videos/vid.mp4 render.type=HUMAN_MASK

# render bounding boxes with track-ids
python scripts/demo.py video.source=assets/videos/vid.mp4 render.type=HUMAN_BBOX

# render a single track id, say 0
python scripts/demo.py video.source=assets/videos/vid.mp4 render.type=TRACKID_0_MESH

# for rendering detected and occluded people
python scripts/demo.py video.source=assets/videos/vid.mp4 render.head_mask=True

You can also visualize the 2D projected keypoints by setting render.show_keypoints=True [TODO].

Track through shot-boundaries

By default, PHALP does not track through shot boundaries. To enable this, please set detect_shots=True.

# for tracking through shot boundaries
python scripts/demo.py video.source=assets/videos/vid.mp4 detect_shots=True

• For debugging purposes, you can set debug=True to disable rich progress bar.

Output .pkl structure

The .pkl file containing tracks, 3D poses, etc. is stored under <video.output_dir>/results, and is a 2-level dictionary:

import joblib
results = joblib.load(<video.output_dir>/results/<video_name>.pkl)
results = {
  # A dictionary for each frame.
  'vid_frame0.jpg': {
    '2d_joints':       List[np.array(90,)],   # 45x 2D joints for each detection
    '3d_joints':       List[np.array(45,3)],  # 45x 3D joints for each detection
    'annotations':     List[Any],             # custom annotations for each detection
    'appe':            List[np.array(4096,)], # appearance features for each detection
    'bbox':            List[[x0 y0 w h]],     # 2D bounding box (top-left corner and dimensions) for each track (detections + ghosts)
    'camera':          List[[tx ty tz]],      # camera translation (wrt image) for each detection
    'camera_bbox':     List[[tx ty tz]],      # camera translation (wrt bbox) for each detection
    'center':          List[[cx cy]],         # 2D center of bbox for each detection
    'class_name':      List[int],             # class ID for each detection (0 for humans)
    'conf':            List[float],           # confidence score for each detection
    'frame_path':      'vid_frame0.jpg',      # Frame identifier
    'loca':            List[np.array(99,)],   # location features for each detection
    'mask':            List[mask],            # RLE-compressed mask for each detection
    'pose':            List[np.array(229,)],  # pose feature (concatenated SMPL params) for each detection
    'scale':           List[float],           # max(width, height) for each detection
    'shot':            int,                   # Shot number
    'size':            List[[imgw imgh]],     # Image dimensions for each detection
    'smpl':            List[Dict_SMPL],       # SMPL parameters for each detection: betas (10), body_pose (23x3x3), global_orient (3x3)
    'tid':             List[int],             # Track ID for each detection
    'time':            int,                   # Frame number
    'tracked_bbox':    List[[x0 y0 w h]],     # 2D bounding box (top-left corner and dimensions) for each detection
    'tracked_ids':     List[int],             # Track ID for each detection
    'tracked_time':    List[int],             # for each detection, time since it was last seen
  },
  'vid_frame1.jpg': {
    ...
  },
  ...
}

Postprocessing pipeline

Coming soon.

Training and Evaluation

Coming soon.

Acknowledgements

Parts of the code are taken or adapted from the following repos:

• deep sort
• SMPL-X
• SMPLify-X
• SPIN
• VIBE
• SMALST
• ProHMR
• TrackEval

Citation

If you find this code useful for your research or the use data generated by our method, please consider citing the following paper:

@inproceedings{rajasegaran2022tracking,
  title={Tracking People by Predicting 3{D} Appearance, Location \& Pose},
  author={Rajasegaran, Jathushan and Pavlakos, Georgios and Kanazawa, Angjoo and Malik, Jitendra},
  booktitle={CVPR},
  year={2022}
}