AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

Degradation-Agnostic Correspondence From Resolution-Asymmetric Stereo

Xihao Chen, Zhiwei Xiong, Zhen Cheng, Jiayong Peng, Yueyi Zhang, Zheng-Jun Zha

CVPR 2022 · Label-efficient and robust learning › Recognition and long-tail · approach: Correspondence and causal representation learning · 10 citations (OpenAlex)

Check your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace photometric consistency with feature-metric consistency from the stereo network's own feature extractorPSMNet unsupervised stereo matching · Inria_SLFD, Middlebury, HCI, and KITTI2015 (BIC, s=4)3PE (%): Inria_SLFD: 8.89 -> 7.70; Middlebury: 6.76 -> 5.78; HCI: 5.13 -> 4.08; KITTI2015: 9.94 -> 8.66Improved
add self-boosting stages that update the feature-metric loss progressively (ablation)PSMNet on Inria_SLFD · s=43PE (%): 12.56 -> 9.22 -> 7.80 -> 7.70Improved
add self-boosting stages that update the feature-metric loss progressively (ablation)PSMNet on Inria_SLFD · s=63PE (%): 21.47 -> 13.92 -> 10.54 -> 9.88Improved
add self-boosting stages that update the feature-metric loss progressively (ablation)PSMNet on Inria_SLFD · s=83PE (%): 27.93 -> 18.47 -> 14.30 -> 13.30Improved
use a symmetric photometric loss instead of an asymmetric photometric loss (ablation)PSMNet on Inria_SLFD · s=2, 4, 6, 83PE (%): 7.28 -> 6.38; 12.56 -> 6.39; 22.72 -> 6.58; 27.93 -> 7.52Improved
use symmetric rather than asymmetric inputs to the feature extractor while keeping the asymmetric loss (ablation)PSMNet on Inria_SLFD · s=2, 4, 6, 83PE (%): 7.28 -> 7.22; 12.56 -> 10.01; 22.72 -> 16.31; 27.93 -> 21.93Improved
use CL features as the feature-metric space (ablation)BaseNet · simulated asymmetric stereo datasets3PE (%): Inria_SLFD BIC: 12.56 -> 12.97; HCI BIC: 5.95 -> 7.80; KITTI2015 BIC: 11.32 -> 17.04Hurt
use autoencoder features as the feature-metric space (ablation)BaseNet · simulated asymmetric stereo datasets3PE (%): Inria_SLFD BIC: 12.56 -> 10.47; HCI BIC: 5.95 -> 5.13; KITTI2015 BIC: 11.32 -> 10.53Mixed
replace PSMNet with iResNet while retaining the proposed feature-metric method (ablation)Ours on simulated datasets · BIC, s=43PE (%): Inria_SLFD: 18.80 -> 9.83; HCI: 18.58 -> 5.83; Middlebury: 10.92 -> 8.39; KITTI2015: 17.82 -> 10.86Improved
use supervised disparity labels instead of unsupervised learning (ablation)BaseNet on Middlebury-trained models · KITTI2015 test, IG, s=43PE (%): 13.14 -> 19.46Hurt
use RCAN super-resolution preprocessingBaseNet · real-world Huawei P30 tele-wide stereovisual disparity qualityNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy