Uncertainty Estimation for Safety-critical Scene Segmentation via Fine-grained Reward Maximization
Authors: Hongzheng Yang, Cheng Chen, Yueyao CHEN, Scheppach, Hon Chi Yip, DOU QI
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The effectiveness of our method is demonstrated on two large safety-critical surgical scene segmentation datasets under two different uncertainty estimation settings. With real-time one forward pass at inference, our method outperforms state-of-the-art methods by a clear margin on all the calibration metrics of uncertainty estimation, while maintaining a high task accuracy for the segmentation results. |
| Researcher Affiliation | Academia | Hongzheng Yang1 , Cheng Chen2 , Yueyao Chen1, Markus Scheppach3, Hon Chi Yip1, Qi Dou1 1The Chinese University of Hong Kong 2Harvard Medical School & Massachusetts General Hospital 3 University Hospital of Augsburg |
| Pseudocode | Yes | Algorithm 1 FGRM algorithm |
| Open Source Code | Yes | Code is available at https://github.com/med-air/FGRM. |
| Open Datasets | Yes | Dataset-1: For LC segmentation dataset, we adopt the public dataset Cholec Seg8K [17], which contains 8,080 laparoscopic cholecystectomy image frames extracted from 17 video clips. Dataset-2: For ESD segmentation dataset, we collected a dataset with 1,203 image frames from 30 endoscopic surgical videos. ... we also provide experimental evaluations on Cityscape [10] dataset for urban scene segmentation in Appendix A.5. |
| Dataset Splits | Yes | For each dataset, we first randomly split 20% data for a held-out testing, and further split 80% of remaining data for training and 20% for validation. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. |
| Software Dependencies | No | In our implementation, we employ an adapted Trans UNet as segmentation backbone. We replace the last softmax layer with a non-negative evidence layer. The evidence layer is implemented by the softplus function. For the base model pre-training, we use the Adam optimizer, with learning rate initialized to 1e-4. |
| Experiment Setup | Yes | For the base model pre-training, we use the Adam optimizer, with learning rate initialized to 1e-4. We totally trained 10 epoches on the training set, with batch size 4. For the maximization of uncertainty estimation reward, we tune the base model to maximize the reward on a held-out validation set. ... The learning rate and batch size was initialized as 1e-4 and 4, respectively. |