TVNet: Temporal Voting Network for Action Localization

This repo holds the codes of paper: "TVNet: Temporal Voting Network for Action Localization".

Paper Introduction

Temporal action localization is a vital task in video understranding. In this paper, we propose a Temporal Voting Network (TVNet) for action localization in untrimmed videos. This incorporates a novel Voting Evidence Module to locate temporal boundaries, more accurately, where temporal contextual evidence is accumulated to predict frame-level probabilities of start and end action boundaries.

Dependencies

• Python == 2.7
• Tensorflow == 1.9.0
• CUDA==10.1.105
• GCC >= 5.4

Note that the PEM code from BMN is implemented in Pytorch==1.1.0 or 1.3.0

Data Preparation

Datasets

Our experiments is based on ActivityNet 1.3 and THUMOS14 datasets.

Feature for THUMOS14

You can download the feature on THUMOS14 at here GooogleDrive.

Place it into a folder named thumos_features inside ./data.

You also need to download the feature for PEM (from BMN) at GooogleDrive. Please put it into a folder named Thumos_feature_hdf5 inside ./TVNet-THUMOS14/data/thumos_features.

If everything goes well, you can get the folder architecture of ./TVNet-THUMOS14/data like this:

data                       
└── thumos_features                    
		├── Thumos_feature_dim_400              
		├── Thumos_feature_hdf5               
		├── features_train.npy 
		└── features_test.npy

Feature for ActivityNet 1.3

You can download the feature on ActivityNet 1.3 at here GoogleCloud. Please put csv_mean_100 directory into ./TVNet-ANET/data/activitynet_feature_cuhk/.

If everything goes well, you can get the folder architecture of ./TVNet-ANET/data like this:

data                        
└── activitynet_feature_cuhk                    
		    └── csv_mean_100

Run all steps

Run all steps on THUMOS14

cd TVNet-THUMOS14

Run the following script with all steps on THUMOS14:

bash do_all.sh

Note: If you use BlueCrystal 4, you can directly run the following script without any dependencies setup.

bash do_all_BC4.sh

Run all steps on ActivityNet 1.3

cd TVNet-ANET
bash do_all.sh  or  bash do_all_BC4.sh

Run steps separately

Take TVNet-THUMOS14 as an example:

cd TVNet-THUMOS14

1. Temporal evaluation module

python TEM_train.py

python TEM_test.py

2. Creat training data for voting evidence module

python VEM_create_windows.py --window_length L --window_stride S

L is the window length and S is the sliding stride. We generate training windows for length 10 with stride 5, and length 5 with stride 2.

3. Voting evidence module

python VEM_train.py --voting_type TYPE --window_length L --window_stride S

python VEM_test.py --voting_type TYPE --window_length L --window_stride S

TYPE should be start or end. We train and test models with window length 10 (stride 5) and window length 5 (stride 2) for start and end separately.

4. Proposal evaluation module from BMN

python PEM_train.py

5. Proposal generation

python proposal_generation.py

6. Post processing and detection

python post_postprocess.py

Results

THUMOS14

ActivityNet 1.3

Reference

This implementation borrows from:

BSN: BSN-Boundary-Sensitive-Network

TEM_train/test.py -- for the TEM module we used in our paper
load_dataset.py -- borrow the part which load data for TEM

BMN: BMN-Boundary-Matching-Network

PEM_train.py -- for the PEM module we used in our paper

G-TAD: Sub-Graph Localization for Temporal Action Detection

post_postprocess.py -- for the multicore process to generate detection

Our main contribution is in:

VEM_create_windows.py -- generate training annotations for Voting Evidence Module (VEM)

VEM_train.py -- train Voting Evidence Module (VEM)

VEM_test.py -- test Voting Evidence Module (VEM)