- This is the implementation of the model proposed in [1].
- Please refer to Caoyichao/STCNet for the repo given by the author of [1]. Note that, by the time I shared my code, the authors of [1] have not made their implementation available. I shared the code for research study only.
- The result (the F-score on testing dataset) I got tally with the one shown in [1], which indicates the correctness of the implementation of this repo.
- I used and modified part of the code in Cadene/pretrained-models.pytorch for the implementation of the backbone (SE-ResNext).
- The data we used is given by [2]. I used and modified part of the code in CMU-CREATE-Lab/deep-smoke-machine for downloading and pre-processing data.
- PyTorch. The version I used was 1.9.0a0+git1fca154. Please note that the stable version 1.9.0 has been available by the time I shared this repo. The version of 1.9.0 should be better.
- Download and pre-process data (videos with 320 by 320 resolutions)
bash data_preprocess.sh 320
- Training and validating (e.g., using GPU 0). The tensorboard log, prediction outputs (on validation data), and the trained model will be saved.
python main.py --gpu 0
- Validating
python main.py --test --mode validation
- Testing
python main.py --test --mode test
- The trained model is saved in trained_model.
- The prediction outputs are saved in pred_out.
- Results on validation dataset:
- Accuracy 0.925699 - Precision 0.923586 - Recall 0.913358 - F-score 0.918443
- Results on testing dataset:
- Accuracy 0.917490 - Precision 0.900443 - Recall 0.876863 - F-score 0.888497
- [1] Y. Cao, Q. Tang, X. Lu, F. Li, and J. Cao, “STCNet: Spatio-Temporal Cross Network for Industrial Smoke Detection,”arXiv preprintarXiv:2011.04863, 2020.
- [2] Y.-C. Hsu, T.-H. K. Huang, T.-Y. Hu, P. Dille, S. Prendi, R. Hoffman, A. Tsuhlares, J. Pachuta, R. Sargent, and I. Nourbakhsh, “Project RISE: Recognizing Industrial Smoke Emissions,” in Proc. of AAAI, 2021.