Causality In Traffic Accident (Under Construction)

Repository for Traffic Accident Benchmark for Causality Recognition (ECCV 2020)

Overview

Data Preparation

Details of dataset construction

Benchmark

Cause and Effect Event Classification

We adopt Temporal Segment Networks (ECCV 2016) from the repository https://github.com/yjxiong/tsn-pytorch

• The default arguments for code are set to train TSN with average consensus function.

python train_classifier.py --consensus_type average
python train_classifier.py --consensus_type linear

Temporal Cause and Effect Event Localization

We adopt three types of baseline methods in our benchmark.

• Single-stage Action Detection

python train_localization.py --architecture_type forward-SST
python train_localization.py --architecture_type backward-SST
python train_localization.py --architecture_type bi-SST
python train_localization.py --architecture_type SSTCN-SST

• Proposal-based Action Detection (Not supported yet)

python train_localization.py --architecture_type naive-conv-R-C3D
python train_localization.py --architecture_type SSTCN-R-C3D

• Action Segmentation

python train_localization.py --architecture_type SSTCN-Segmentation
python train_localization.py --architecture_type MSTCN-Segmentation