Any-Stereo: Arbitrary Scale Disparity Estimation for Iterative Stereo Matching
Authors: Zhaohuai Liang, Changhe Li
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments Datasets & Evaluation Metrics Datasets The Secene Flow dataset (Mayer et al. 2016), KITTI dataset (Geiger, Lenz, and Urtasun 2012; Menze and Geiger 2015) and Middlebury dataset (Scharstein et al. 2014) are used. [...] Table 1: Quantitative evaluation on KITTI 2015 and KITTI 2012. |
| Researcher Affiliation | Academia | Zhaohuai Liang 1, Changhe Li 2* 1 School of Automation, China University of Geosciences, Wuhan 430074, China 2 School of Artificial Intelligence, Anhui University of Science & Technology, Hefei 232001, China |
| Pseudocode | No | The paper describes the proposed methods and their components in detail but does not provide pseudocode or a clearly labeled algorithm block. |
| Open Source Code | Yes | The code is available at https://github.com/Zhaohuai-L/Any-Stereo. |
| Open Datasets | Yes | Datasets The Secene Flow dataset (Mayer et al. 2016), KITTI dataset (Geiger, Lenz, and Urtasun 2012; Menze and Geiger 2015) and Middlebury dataset (Scharstein et al. 2014) are used. |
| Dataset Splits | Yes | KITTI12 consists of 194 training pairs and 195 testing pairs, and KITTI15 consists of 200 training pairs and 200 testing pairs. [...] All models are trained on Scene Flow training sets, and evaluated on the testing sets with a fixed scale to full resolution. |
| Hardware Specification | Yes | we perform our experiments on two NVIDIA A40 GPUs. |
| Software Dependencies | No | The paper mentions 'Pytorch' but does not specify a version number or other software dependencies with their versions. |
| Experiment Setup | Yes | For all training, we use the Adam W (Kingma and Ba 2014) optimizer with a one-cycle learning rate schedule and clip gradients to [-1,1]. For the ISU, we set window size p as 5. [...] The model for ablation is trained on Scene Flow for 100k steps with a batch size 6. The final model is pretrained on Scene Flow for 200k steps with a batch size of 8, and then finetuned on KITTI and Middlebury. All experiments are run with 22 update iterations during training, 32 during evaluation. |