Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Accident Anticipation via Temporal Occurrence Prediction
Authors: Tianhao Zhao, Yiyang Zou, Zihao Mao, Peilun Xiao, Yulin Huang, Hongda Yang, Yuxuan Li, Tracy Li, Guobin Wu, Yutian Lin
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that our method achieves superior performance in both recall and TTA under realistic FAR constraints. |
| Researcher Affiliation | Collaboration | 1School of Computer Science, Wuhan University 2Zhongguancun Academy, Beijing, China 3Didi Chuxing EMAIL |
| Pseudocode | No | The paper describes the methodology in prose and uses diagrams (Figure 2) but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Project page: https://happytianhao.github.io/TOP/ and Our code, model checkpoints, and detailed instructions for reproducing the main results are provided on the project page: https://happytianhao.github.io/ TOP/. |
| Open Datasets | Yes | All datasets used in this work (CAP [9], DADA [8], and MM-AU [33]) are publicly available. |
| Dataset Splits | No | We selected approximately 20% of the data as the test set; for evaluation, we extract clips from the first frame to the anomaly appear frame as negative samples to compute the false alarm rate (FAR), and clips from anomaly appear to accident occur as positive samples to assess anticipation recall and Time-to-Accident (TTA). This only specifies the test set size as "approximately 20%" and describes how samples are extracted for evaluation, but does not provide explicit training/validation split percentages or sample counts. |
| Hardware Specification | Yes | We optimize the model using SGD with a batch size of 64 on 8 NVIDIA 4090 GPUs. |
| Software Dependencies | No | The paper mentions using SGD as an optimizer but does not specify software libraries or frameworks with version numbers (e.g., PyTorch, TensorFlow, Python) used for implementation. |
| Experiment Setup | Yes | Each input snippet consists of S = 5 consecutive frames, which are fed into a snippet-level encoder (Slow Only [34], initialized with Image Net pre-trained weights) to extract spatiotemporal features. The model then predicts a sequence of accident scores of length T = 20, corresponding to future time steps from 0.1s to 2.0s ahead. To decode these scores, we employ a Transformer-based temporal decoder with 2 layers and cosine positional encodings as queries for each future horizon. We optimize the model using SGD with a batch size of 64 on 8 NVIDIA 4090 GPUs. The binary cross-entropy loss is weighted with w+ = 10 for positive samples, and the initial learning rate is set to 0.01, decayed to 10% of its value every 20 epochs over 50 total epochs. |