You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
@inproceedings{kukleva2019unsupervised,
title={Unsupervised learning of action classes with continuous temporal embedding},
author={Kukleva, Anna and Kuehne, Hilde and Sener, Fadime and Gall, Jurgen},
booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR'19)},
year={2019}
}
Pipeline for one activity class. Figure 1 in the paper.
Proposed pipeline for unsupervised learning with unknown activity classes. Figure 2 in the paper.
Visualization of embeddings via t-SNE on the 50Salads dataset
Each frame is color coded a) with the corresponding ground truth subaction label, b) with K assigned subaction labels after clustering as the second step in Fig.1 in our main paper, c)with the predicted labels after the decoding stage. The optimization of our network is performed with respect to relative timestep of each frame. In d) we show the respective relative time label in the continuous temporal embedding assigned to each frame feature. The color bar depicts that bright blue corresponds to 0 (startof the video) and pink to 1 (end of the video).
The number of subactions K
Breakfast dataset
Activity class name
# subactions (K)
Coffe
7
Cereals
5
Tea
7
Milk
5
Juice
8
Sandwich
9
Scrambledegg
12
Friedegg
9
Salat
8
Pancake
14
YouTube Instractions dataset
Activity class name
# subactions (K)
Changing tire
11
Making cofee
10
CPR
7
Jump car
12
Repot plant
8
Qualitative results
Breakfast dataset. The order of subactions: SIL, take bowl, pour cereals, pour milk, stir cereals, SIL
Breakfast dataset. The order of subactions: SIL, take cup, add teabag, pour water, SIL
Breakfast dataset. The order of subactions: SIL, spoon powder, pour milk, stir milk, SIL
Breakfast dataset. The order of subactions: SIL, take knife, cut orange, squeeze orange, pour juice, squeeze orange, pour juice, squeeze orange, pour juice, squeeze orange, pour juice, SIL
Breakfast dataset. The order of subactions: SIL, cut bun, smear butter, put toppingOnTop, SIL
50Salads dataset. The order of subactions: start, cut, place, cut, place, cut, place, cut, place, null, null, add oil, add pepper, mix dressing, end
50Salads dataset. The order of subactions: start, cut, place, cut, place, cut, place, peel cucumber, cut, place, mix ingredients, add oil, null, add pepper, null, mix dressing, serve salad onto plate, add dressing, end
About
Unsupervised learning of action classes with continuous temporal embedding (CVPR'19)