Revisiting Non-Parametric Matching Cost Volumes for Robust and Generalizable Stereo Matching

Authors: Kelvin Cheng, Tianfu Wu, Christopher Healey

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments, the proposed method is tested in the Scene Flow dataset, the KITTI2015 dataset, and the Middlebury dataset. It significantly improves the adversarial robustness, while retaining accuracy performance comparable to state-of-the-art methods.
Researcher Affiliation Academia Kelvin Cheng , Tianfu Wu and Christopher Healey CS and ECE at NC State University, Raleigh, NC 27695 {kbcheng, twu19, healey}@ncsu.edu
Pseudocode No The paper does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our code and pretrained models are released at this Github Repo.
Open Datasets Yes We evaluate our method on the Scene Flow [5] and KITTI2015 [6] datasets. We also test pretrained models on the KITTI2012 [15] and the Middlebury [7] dataset at quarter resolution.
Dataset Splits Yes The Scene Flow dataset is a large-scale synthetic dataset that contains 35, 454 training images and 4, 370 test images at the resolution of 540 × 960. The KITTI2015 dataset is a real-world dataset of driving scenes, which contains 200 training images and 200 test images at the resolution of 375 × 1242.
Hardware Specification No The main text of the paper does not specify any particular GPU, CPU, or other hardware details used for running experiments. Appendix D is referenced for compute details, but not provided in the given text.
Software Dependencies No The paper mentions software components like 'smooth L1 loss' and 'sigmoid function', but does not provide specific version numbers for any software libraries or frameworks used.
Experiment Setup Yes In our experiments, we set = 0.06 or 0.03... = 0.01 and T = 20. We use local windows with sizes from k1 to k2 (e.g. k1 = 3, k2 = 11 in our experiments)... (e.g., 0.5, 0.7, and 1 are used for the 3-stack Hourglass module in our experiments).