Object tracklets for VidOR train/validation/test set

• you can download the object tracklets for VidOR dataset here
• the object bounding boxes and categories are obtained by MEGA
• the tracking algorithm we used is deepSORT
• these tracklets (these .npy files) only contain tracklet postions and object categories.
• the appearance features (e.g., RoI pooled feature) for each bbox are not released due their large capacity.
• please refer to format_demo.py for detailed format.

for VidOR train set

VidORtrain_freq1_m60s0.3_part01 ~ VidORtrain_freq1_m60s0.3_part14

each part contains 500 videos (500 .npy files)

for VidOR validation set

VidORval_freq1_m60s0.3, which contains 835 videos

for VidOR test set

VidORtest_freq1_m60s0.3, which contains 2165 videos

Explanation of file name and parameters

we explain some parameters in the file names (e.g., VidORtrain_freq1_m60s0.3_part01/0000_2401075277.npy)

• freq1: the sample rate is 1, i.e., we run MEGA and deepSORT on each frame of the video (despite of the large redundancy )

• m60: the parameter max_age in deepSORT tracker, which controls ``maximum number of missed misses before a track is deleted''. m60 means that we allow 60 missed frames for a tracklet

• s0.3: the score threshold is deepSORT, the bounding box with confidence lower than 0.3 will be deleted and will not be considered for tracking.

• 0000_2401075277: 0000 is the group id and 2401075277 is the video id, i.e., this "0000_2401075277.npy" corresponds to "0000/2401075277.mp4" in VidOR train set.

Object tracklets for VidVRD train/test set

• VidVRD_train_every1frames: 800 .npy files
• VidVRD_test_every1frames: 200 .npy files
• these .npy files have the same format as that in VidOR. Refer to format_demo.py
• The parameters we used in deepSORT for VidVRD: m30s0.3