Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

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, for instructions.

Training and Inference

Please see for details.


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.


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 for instructions on how to do this.


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


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


  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}}


Long-Term Feature Banks for Detailed Video Understanding




No releases published


No packages published