UM

A PyTorch implementation of UM based on AAAI 2021 paper Weakly-supervised Temporal Action Localization by Uncertainty Modeling.

Requirements

conda install pytorch=1.10.0 torchvision cudatoolkit=11.3 -c pytorch

MMAction2

pip install git+https://github.com/open-mmlab/mim.git
mim install mmaction2

Dataset

THUMOS 14 and ActivityNet datasets are used in this repo, you could download these datasets from official websites. The I3D features of THUMOS 14 dataset can be downloaded from Google Drive, I3D features of ActivityNet 1.2 dataset can be downloaded from OneDrive , I3D features of ActivityNet 1.3 dataset can be downloaded from Google Drive. The data directory structure is shown as follows:

├── thumos14                                    |  ├── activitynet
  ├── features                                  |    ├── features_1.2
      ├── val                                   |        ├── train 
          ├── flow                              |            ├── flow    
              ├── video_validation_0000051.npy  |                ├── v___dXUJsj3yo.npy
              └── ...                           |                └── ...
          ├── rgb (same structure as flow)      |            ├── rgb
      ├── test                                  |                ├── v___dXUJsj3yo.npy
          ├── flow                              |                └── ...
              ├── video_test_0000004.npy        |        ├── val (same structure as tain)
              └── ...                           |    ├── features_1.3 (same structure as features_1.2)
          ├── rgb (same structure as flow)      |    ├── videos
  ├── videos                                    |        ├── train
      ├── val                                   |            ├── v___c8enCfzqw.mp4
          ├── video_validation_0000051.mp4      |            └──... 
          └──...                                |         ├── val           
      ├── test                                  |            ├── v__1vYKA7mNLI.mp4
          ├──video_test_0000004.mp4             |            └──...   
          └──...                                | annotations_1.2.json
  annotations.json                              | annotations_1.3.json

Usage

You can easily train and test the model by running the script below. If you want to try other options, please refer to utils.py.

Train Model

python train.py --data_name activitynet1.2 --num_segments 50 --seed 0 --scale 16

Test Model

python test.py --model_file --data_name thumos14 --model_file result/thumos14_model.pth

Benchmarks

The models are trained on one NVIDIA GeForce GTX 1080 Ti GPU (11G). All the hyper-parameters are the default values according to the papers.

THUMOS14

Method	THUMOS14								Download
	mAP@0.1	mAP@0.2	mAP@0.3	mAP@0.4	mAP@0.5	mAP@0.6	mAP@0.7	mAP@AVG
	Ours	60.3	54.3	45.7	37.2	27.8	18.2	9.2		36.1	kb79
Official	67.5	61.2	52.3	43.4	33.7	22.9	12.1	41.9	-

mAP@AVG is the average mAP under the thresholds 0.1:0.1:0.7.

ActivityNet

Method	ActivityNet 1.2				ActivityNet 1.3				Download
	mAP@0.5	mAP@0.75	mAP@0.95	mAP@AVG	mAP@0.5	mAP@0.75	mAP@0.95	mAP@AVG
	Ours	1.7	0.5	0.0	0.7	0.1	0.1	0.0		0.1	wexe
Official	41.2	25.6	6.0	25.9	37.0	23.9	5.7	23.7	-

mAP@AVG is the average mAP under the thresholds 0.5:0.05:0.95.

Reference

This repo is built upon the repo WTAL-Uncertainty-Modeling.

leftthomas/UM

UM