Skip to content

Latest commit

 

History

History
30 lines (25 loc) · 2.14 KB

DATASET.md

File metadata and controls

30 lines (25 loc) · 2.14 KB
Raw

PHALP Dataset

We ran PHALP on the kinetics 400 and AVA to collect about 1.5 million human trajectories. Eack .pkl is a single tracklet. Naming convention is <DATASET>_<VIDEO_ID>_<KEY_FRAME>_<TRACK_ID>_<TRACK_LENGTH> . For example ava-train_053oq2xB3oU_000150_3_128.pkl means, this tracklet is from the ava-train dataset, from the video 053oq2xB3oU, key frame is 000150 (will have annotations), with the track id 3 and the track length is 128 frames. Each tracklet contains the following information:

import joblib

track = joblib.load("data/ava_train/ava-train_053oq2xB3oU_000150_3_128.pkl")

# Tracklet information
>>> track.keys()
dict_keys(['fid', 'pose_shape', '3d_joints', 'camera', 'camera_bbox', 'action_label_gt', 'action_label_psudo', 'has_gt', 'has_detection', 'apperance_index', 'apperance_dict', 'frame_name', 'frame_size', 'frame_bbox', 'frame_conf'])
  • fid: counter for the frames in the tracklet # (128, 1, 1)
  • pose_shape: SMPL pose, shape, global orient, and camera parameters # (128, 1, 229)
  • 3d_joints: 3D joints in camera coordinates # (128, 1, 45, 3)
  • camera: camera parameters # (128, 1, 3)
  • camera_bbox: camera with wrt bounding box # (128, 1, 3)
  • action_label_gt: ground truth action label # (128, 1, 80)
  • action_label_psudo: pseudo ground truth action label # <class 'dict'>, contains labels according to apperance_index
  • has_gt: whether the tracklet has ground truth annotations # (128, 1, 1), 2 means gt is available, 1 means pseudo gt is available, 0 means no gt is available
  • has_detection: whether the tracklet has detections # (128, 1, 1)
  • apperance_index: index of the apperance features and pesudo labels of the dictionary # (128, 1, 1)
  • apperance_dict: dictionary of apperance features # <class 'dict'>, contains apperance features from Hiera indexed by apperance_index
  • frame_name: frame names # list of strings
  • frame_size: frame size # (128, 1, 2)
  • frame_bbox: frame bounding box # (128, 1, 4)
  • frame_conf: frame confidence # (128, 1, 1)

Please see the phalp_action_dataset.py dataset file for more details.