Research map / Detection and segmentation / Open-set and few-shot detection

Hierarchical Object-Aware Dual-Level Contrastive Learning for Domain Generalized Stereo Matching

NeurIPS 2024 · Detection and segmentation › Open-set and few-shot detection · approach: Domain-adaptive object detection · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HODC, a plug-in hierarchical object-aware intra- and inter-scale dual-level contrastive loss that improves synthetic-to-real generalization of stereo matching networks trained only on SceneFlow.

Problem

Stereo matching networks trained on synthetic data generalize poorly to unseen realistic domains because they rely on superficial appearance cues rather than semantic and structural features.

Approach

Segment object index maps into rectangular grids at multiple scales to form object-aware regional representations, warp left features to the right using ground-truth disparity, and apply intra- and inter-scale InfoNCE contrastive losses with hard negative selection, added to the smooth-L1 loss without architecture changes.

Key results

• HODC-PSMNet reduces KITTI-2015 D1 error from 16.3 (PSMNet) to 4.7 and ETH3D from 23.8 to 5.4 when trained only on SceneFlow.
• HODC-IGEV achieves 4.5 on KITTI-2015, 3.8 on KITTI-2012, and 2.7 on ETH3D, the best ETH3D result in Table 1.
• HODC improves GwcNet on Middlebury half from 34.2 to 8.4 and CFNet on KITTI-2012 from 4.7 to 3.8, outperforming prior domain generalization methods in most settings.

Measured effects reported in the paper

Surprising findings

• Lower similarity between hard negatives, rather than feature consistency between matching pixels, appears more predictive of generalization, contrary to the claim in prior FC-Stereo work.
• Segmenting with fewer grids (larger scale) benefits some datasets, and the method remains effective even without object index priors (whole-image segmentation).

Limitations

• Direct integration into existing architectures overlooks the role of network architecture in domain generalization.
• Requires object index maps (provided by SceneFlow or a pretrained segmentation model) and ground-truth disparity for correspondence.
• Training adds computational and memory overhead from contrastive pairs over multiple scales.

Experimental setup

Datasets: SceneFlow KITTI-2012 KITTI-2015 Middlebury ETH3D DrivingStereo

Benchmarks: KITTI-2012 KITTI-2015 Middlebury ETH3D DrivingStereo

Metrics: D1 outlier percentage

Baselines: PSMNet GwcNet CFNet IGEV STTR DSMNet FC-GANet PCWNet CREStereo++ MS-Net ITSA GraftNet

Builds on: FC-GANet InfoNCE smooth L1 loss domain normalization

Closest papers by meaning

• Revisiting Domain Generalized Stereo Matching Networks From a Feature Consistency Perspective (CVPR 2022 · similarity 0.93)
• Domain Generalized Stereo Matching via Hierarchical Visual Transformation (CVPR 2023 · similarity 0.91)
• Robust Synthetic-to-Real Transfer for Stereo Matching (CVPR 2024 · similarity 0.89)
• All-in-One: Transferring Vision Foundation Models into Stereo Matching (AAAI 2025 · similarity 0.88)
• Stereo Anywhere: Robust Zero-Shot Deep Stereo Matching Even Where Either Stereo or Mono Fail (CVPR 2025 · similarity 0.87)
• Stereo Risk: A Continuous Modeling Approach to Stereo Matching (ICML 2024 · similarity 0.87)