Evidential Learning-based Certainty Estimation for Robust Dense Feature Matching
Authors: Lile Cai, Chuan Sheng Foo, Xun Xu, ZAIWANG GU, Jun Cheng, xulei yang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method on a wide range of benchmarks and show that our method leads to improved robustness against common corruptions and adversarial attacks, achieving up to 10.1% improvement under severe corruptions. We evaluate our method on both corrupted data and clean data, and show that our method leads to improved robustness against common corruptions and adversarial attacks without sacrificing the performance on clean data. |
| Researcher Affiliation | Academia | Institute for Infocomm Research (I2R), A*STAR, Singapore EMAIL |
| Pseudocode | No | The paper describes the methodology using prose and mathematical equations without providing any explicit pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions following the training setup in existing Ro Ma and DKM repositories: 'We follow the training setup in Ro Ma s repo (https://github.com/Parskatt/Ro Ma).' and 'We follow the training setup in DKM s repo (https://github.com/Parskatt/DKM).' However, it does not explicitly state that the source code for the proposed evidential learning framework is made publicly available. |
| Open Datasets | Yes | Mega Depth-1500 (Sun et al., 2021) is a popular outdoor benchmark...For indoor benchmark, we adopt the popular Scan Net-1500 (Sarlin et al., 2020)...The Mega Depth dataset (Li & Snavely, 2018)...The Scan Net dataset (Dai et al., 2017a)... |
| Dataset Splits | Yes | For outdoor geometry estimation, the model is trained on the Mega Depth dataset (Li & Snavely, 2018) using the same training and test splits as in Ro Ma. The training and test split of the dataset contains 1201 and 312 scenes, respectively. (for Scan Net dataset) |
| Hardware Specification | Yes | Training is conducted on 8 A40 GPUs with batch size 32, input size 560 560 and 250k iterations. Inference is conducted on 1 A40 GPU with batch size 1 (1 pair of images) and input size 672 672. Training takes approximately 5 days on a server with 8 A40 GPUs. The model is trained on a server with 8 A5000 GPUs with batch size 16, image resolution 540 720 and training iteration 500k. |
| Software Dependencies | No | The paper mentions using specific models like DINOv2 and VGG19, and the AdamW optimizer, but does not provide specific version numbers for software libraries or frameworks such as Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | We follow the training setup of Ro Ma (Edstedt et al., 2024), where the batch size is 32, encoder learning rate is 10^-4, decoder learning rate is 5e-6, and the model is trained for 250,000 iterations. The training image resolution is 560 x 560. For balanced sampling, we use score threshold 0.05 and sample 5,000 matches. |