renjie-liang / VMRFrame

A frame for Video Moment Retrival

SeqPAN by PyTroch

A SeqPAN model was reproduced by PyTorch for the paper "Parallel Attention Network with Sequence Matching for Video Grounding".

The original TensorFLow implementation is in https://github.com/IsaacChanghau/SeqPAN.

Performance

version	R1@.03	R1@0.5	R1@0.7	mIoU
TensorFlow(paper)	61.65	45.50	28.37	45.11
PyTorch(ours)

Prepareation

Environment

conda create --name tmp python=3.9
pip install -r requirements.txt

Video Feature

To download the video feature, please refer to https://github.com/IsaacChanghau/SeqPAN. It is possible to save the video feature in any preferred location.

Adjust "paths" value within the ./config/anet/SeqPAN.yaml.

Quick Start

# train
CUDA_VISIBLE_DEVICES=0 python main.py --config ./config/anet/SeqPAN.yaml
# test
CUDA_VISIBLE_DEVICES=0 python main.py --config ./config/anet/SeqPAN.yaml --eval
# debug
CUDA_VISIBLE_DEVICES=0 python main.py --config ./config/anet/SeqPAN.yaml --debug