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

Main contributions of the paper

• We introduce a traffic accident analysis benchmark, denoted by CTA, which contains temporal intervals of a cause and an effect in each accident and their semantic labels provided by the crash avoidance research.
• We construct the dataset based on the semantic taxonomy in the crash avoidance research, which makes the distribution of the benchmark coherent to the semantic taxonomy and the real-world statistics.
• We analyze traffic accident tasks by comparing multiple algorithms for temporal cause and effect event localization.

You can download the dataset in the below link Details of dataset

We adopt Temporal Segment Networks (ECCV 2016) in our benchmark.

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

python train_classifier.py --consensus_type average --random_seed 17
python train_classifier.py --consensus_type linear --random_seed 3

• The performance of classification models with above arguments is shown in below.

We adopt three types of baseline methods (single-stage action detection, proposal-based action detection and action segmentation) in our benchmark. Our implementation of methods is based on below three works.

SST: Single-Stream Temporal Action Proposals, CVPR 17 R-C3D: Region Convolutional 3D Network for Temporal Activity Detection, ICCV 2017 MS-TCN: Multi-Stage Temporal Convolutional Network for Action Segmentation, CVPR 19

• 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 --num_layers 10 --num_epochs 100

For single-stage detection, we adopt SST. We use K = 128 for the size of the hidden dimension for gated recurrent units (GRU). To change the proposed method into a single-stage detection method, we simply change the class prediction layer to have three classes background, cause and effect—and substitute binary cross-entropy loss function into cross-entropy loss function. We use 64 anchor boxes with temporal scales [1 · δ, 2 · δ, · · · , K · δ] in seconds, where δ = 0.32 seconds and K = 64.

Note that the performances of backward-SST, Bi-SST and SSTCN-SST except forward-SST are better than those in the paper.

• Action Segmentation

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

• 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

@inproceedings{you2020CTA,
    title     = "{Traffic Accident Benchmark for Causality Recognition}",
    author    = {You, Tackgeun and Han, Bohyung},
    booktitle = {ECCV},
    year      = {2020}
}