Stereo Risk: A Continuous Modeling Approach to Stereo Matching
Authors: Ce Liu, Suryansh Kumar, Shuhang Gu, Radu Timofte, Yao Yao, Luc Van Gool
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | A comprehensive analysis demonstrates our method s theoretical soundness and superior performance over the state-of-the-art methods across various benchmark datasets, including KITTI 2012, KITTI 2015, ETH3D, Scene Flow, and Middlebury 2014. |
| Researcher Affiliation | Academia | 1Nanjing University, China. 2ETH Z urich, Switzerland. 3VCCM, Texas A&M University, USA. 4UESTC, China. 5University of W urzburg, Germany. 6KU Leuven, Belgium. 7INSAIT, Bulgaria. |
| Pseudocode | Yes | Algorithm 1 Forward Prediction input τ > 0, σ > 0, d = [d1, ..., d N], d1 < d2 < < d N, and pm = [pm 1 , ..., pm N] dl d1 dr d N g τ + 1 while |g| > τ do dm (dl + dr)/2.0 g P i pm i Sign(dm di)(1 exp |dm di| σ ) if g > 0 then end if end while output dm |
| Open Source Code | No | The paper does not explicitly state that its source code is open-source or provide a link to a code repository. |
| Open Datasets | Yes | Datasets. We perform experiments on four datasets namely Scene Flow (Mayer et al., 2016), KITTI 2012 & 2015 (Geiger et al., 2012; Menze & Geiger, 2015), Middlebury 2014 (Scharstein & Szeliski, 2002), and ETH 3D (Sch ops et al., 2017). |
| Dataset Splits | No | The paper specifies training and test splits (e.g., 'Scene Flow is a synthetic dataset containing 35,454 image pairs for training, and 4,370 image pairs for test.') but does not explicitly provide details for a separate validation split. |
| Hardware Specification | Yes | The software is evaluated on a machine with Ge Force-RTX-3090 GPU. |
| Software Dependencies | Yes | We implement our method in Py Torch 2.0.1 (Python 3.11.2) with CUDA 11.8. |
| Experiment Setup | Yes | We use Adam W optimizer (Loshchilov & Hutter, 2019) with weight decay 10 5. The learning rate decreases from 2 10 4 to 2 10 8 according to the one cycle learning rate policy. We train the network for 2 105 iterations. The images are randomly cropped to 320 736. |