Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


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


This repository contains tensorflow implementation for CTAP: Complementary Temporal Action Proposal Generation in ECCV 2018.


The framework of CTAP is shown in the figure below:

This repository consists of three parts:

  • code for Temporal convolutional Adjustment and Ranking (TAR) network which uses temporal conv layers to aggregate the unit-level features ( in the folder TAR)
  • code for Proposal-level Actionness Trustworthiness Estimator (PATE) classifier ( in the folder PATE)
  • code for complementary filetering ( in the folder complementary_filtering)

Feature Download

We provide both unit-level features for TAG scores prediction and sliding windows features for TAR on THUMOS-14 dataset.

Note: validation set is used for training, as the training set for THUMOS-14 does not contain untrimmed videos.

For unit-level features (unit length = 16), the appearance features can be downloaded at here: val set, test set; the denseflow features can be downloaded here: val set, test set.

For unit-level features (unit length = 6), the appearance and denseflow features for validation and test sets are available in google drive.

TAR Train & Evaluation


Download the unit level featurs, and edit the feature path in TAR/, and then just run python The in TAR/test_results folder should be applied on the output test result file. After post processing, the pkl file can be evaluated by the eval program, which can be found in here.


Currently there are three sets of action proposals for testing with different purposes. To test different proposals, please modify TAR/ file in lines 181-185 respectively.

line 185: test_clip_path test_swin.txt tag_proposals_round6.txt test_swin_unit6_sample4.txt
Evaluated proposals sliding windows with unit length of 16 TAG proposals sliding windows with unit length of 6

When line 185 is test_swin.txt, line 181-184 contents do not need modification. When line 185 becomes tag_proposals_round6.txt or test_swin_unit6_sample4.txt, please modify line 181-184 as the feature folders downloaded and unzipped from the link of unit-level features (unit length = 6) provided on the above. Besides, modify line 14 as unit_size=6.0.

PATE Train & Evaluation

To train PATE, go to PATE folder and then run python The in PATE/test_results folder should be applied, which takes two inputs, one is the PATE output .pkl and the other is TAR output .pkl, these two pkl files should be from the same sliding window file.

Complementary Filtering

The complementary_filter/ combines the results from sliding window and TAG proposals together based on PATE scores, the output is the final proposals. It takes two files as input, one is the sliding window proposals (generated from PATE/test_results/, and the other is TAG proposals (generated from TAR/test_results/

Reproduce the results

The final proposal list on THUMOS-14 generated from CTAP is here, using this eval tool to reproduce our results.


If you find the repository is useful for your research, please consider citing the following work:

  title={CTAP: Complementary Temporal Action Proposal Generation},
  author={Gao*, Jiyang and Chen*, Kan and Nevatia, Ram},


ECCV2018 CTAP: Complementary Temporal Action Proposal Generation







No releases published


No packages published