Long-Term Feature Banks for Detailed Video Understanding

Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krähenbühl, Ross Girshick
In CVPR 2019. [Paper]

This is a Caffe2 based implementation for our CVPR 2019 paper on Long-Term Feature Banks (LFB). LFB provides supportive information extracted over the entire span of a video, to augment state-of-the-art video models that otherwise would only view short clips of 2-5 seconds. Our experiments demonstrate that augmenting 3D CNNs with an LFB yields state-of-the-art results on AVA, EPIC-Kitchens, and Charades.

Data Preparation and Installation

Please see DATASET.md, INSTALL.md for instructions.

Training and Inference

Please see GETTING_STARTED.md for details.

Results

The following documents a collection of models trained with this repository. Links to the trained models and output log files are provided. The performance is evaluated on the validation set.

Note that all models here are not the original models used in paper, but reproduced by this code base. The reproduced performance reported here is very close to (or slightly better than) what's reported in paper, but not exactly the same due to the stochastic nature of training.

AVA

config	backbone	method	mAP	model id	model
ava_r50_baseline	R50-I3D-NL	3D CNN	22.2	102760666	model
ava_r50_lfb_avg	R50-I3D-NL	LFB-Avg	23.3	103505104	model, lfb model
ava_r50_lfb_max	R50-I3D-NL	LFB-Max	23.9	103505159	model, lfb model
ava_r50_lfb_nl	R50-I3D-NL	LFB-NL-2L	25.8	102824705	model, lfb model
ava_r50_lfb_nl_3l	R50-I3D-NL	LFB-NL-3L	25.9	106403526	model, lfb model
ava_r101_baseline	R101-I3D-NL	3D CNN	23.2	102760714	model
ava_r101_lfb_nl_3l	R101-I3D-NL	LFB-NL-3L	26.9 (multi-crop: 27.7)	105206523	model, lfb model

EPIC Kitchens Verb

config	backbone	method	top1	top5	model id	model
epic_verb_r50_baseline	R50-I3D-NL	3D CNN	50.7	81.1	103704809	model
epic_verb_r50_lfb_avg	R50-I3D-NL	LFB-Avg	52.9	82.5	103777391	model, lfb model
epic_verb_r50_lfb_max	R50-I3D-NL	LFB-Max	53.3	81.0	103777432	model, lfb model
epic_verb_r50_lfb_nl	R50-I3D-NL	LFB-NL	52.3	81.8	103777046	model, lfb model

EPIC Kitchens Noun

config	backbone	method	top1	top5	model id	model
epic_noun_r50_baseline	R50-I3D-NL	3D CNN	26.2	51.0	104421642	model
epic_noun_r50_lfb_avg	R50-I3D-NL	LFB-Avg	29.1	56.3	103875866	model
epic_noun_r50_lfb_max	R50-I3D-NL	LFB-Max	32.0	56.5	103875899	model
epic_noun_r50_lfb_nl	R50-I3D-NL	LFB-NL	29.5	55.4	103706990	model

EPIC Kitchens Action

config	backbone	method	top1	top5
epic_verb_r50_baseline & epic_noun_r50_baseline	R50-I3D-NL	3D CNN	19.4	38.1
epic_verb_r50_lfb_avg & epic_noun_r50_lfb_avg	R50-I3D-NL	LFB-Avg	21.2	41.3
epic_verb_r50_lfb_max & epic_noun_r50_lfb_max	R50-I3D-NL	LFB-Max	22.9	41.2
epic_verb_r50_lfb_nl & epic_noun_r50_lfb_nl	R50-I3D-NL	LFB-NL	21.8	40.5

Note: To make action predictions, we combine a verb model and a noun model, as opposed to training a separate action model. Performance in this table is computed using the verb/noun models from the tables above. Please see GETTING_STARTED.md for instructions on how to do this.

Charades

config	backbone	method	mAP	model id	model
charades_r50_baseline	R50-I3D-NL	3D CNN	38.3	102766107	model
charades_r50_lfb_avg	R50-I3D-NL	LFB-Avg	38.4	102999065	model, lfb model
charades_r50_lfb_max	R50-I3D-NL	LFB-Max	38.6	102999121	model, lfb model
charades_r50_lfb_nl	R50-I3D-NL	LFB-NL	40.3	100866795	model, lfb model
charades_r101_baseline	R101-I3D-NL	3D CNN	40.4	103560426	model
charades_r101_lfb_avg	R101-I3D-NL	LFB-Avg	40.8	103676713	model, lfb model
charades_r101_lfb_max	R101-I3D-NL	LFB-Max	41.0	103676788	model, lfb model
charades_r101_lfb_nl	R101-I3D-NL	LFB-NL	42.5	103641815	model, lfb model

License

Video-long-term-feature-banks is Apache 2.0 licensed, as found in the LICENSE file.

Citation

@inproceedings{lfb2019,
  Author    = {Chao-Yuan Wu and Christoph Feichtenhofer and Haoqi Fan
               and Kaiming He and Philipp Kr\"{a}henb\"{u}hl and
               Ross Girshick},
  Title     = {{Long-Term Feature Banks for Detailed Video Understanding}},
  Booktitle = {{CVPR}},
  Year      = {2019}}