Research map / Vision-language and multimodal / Cross-modal learning
ContextDesc: Local Descriptor Augmentation With Cross-Modality Context
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 changed | Where | Measured | Effect |
|---|---|---|---|
| augment GeoDesc with visual regional and geometric keypoint context | GeoDesc · HPSequences | Recall: 59.46 -> 66.55; 71.24 -> 75.52 | Helped |
| use regional visual features instead of a single global feature (ablation) | GeoDesc visual context · HPSequences | Recall: 59.11 -> 63.64; 71.02 -> 73.37 | Helped |
| add context normalization when encoding regional visual features (ablation) | regional visual context · HPSequences | Recall: 63.64 -> 63.98; 73.37 -> 73.63 | Helped |
| use one global visual feature (ablation) | GeoDesc visual context · HPSequences | Recall: 59.46 -> 59.11; 71.24 -> 71.02 | Hurt |
| use PointNet for 2D geometric context (ablation) | GeoDesc geometric context · HPSequences | Recall: 59.46 -> 59.61; 71.24 -> 70.96 | Hurt |
| add matchability to pre-activated CN and xy input (ablation) | geometric context encoder · HPSequences | Recall: 61.67 -> 62.82; 72.63 -> 73.40 | Mixed |
| use original CN residual units rather than pre-activation (ablation) | geometric context encoder with xy and matchability · HPSequences | Recall: 59.94 -> 62.82; 71.25 -> 73.40 | Hurt |
| replace ContextDesc+ with the proposed temperature-scaled N-pair loss for the local base model (ablation) | local descriptor without context · HPatches | mAP: 91.1 -> 90.2; 59.1 -> 59.2; 74.9 -> 76.0 | Mixed |
| train the local descriptor from scratch while keeping the regional model frozen (ablation) | ContextDesc+ · HPSequences | Recall: 67.14 to 67.53; 76.42 to 77.20 | Helped |
| jointly optimize the regional model with the descriptor (ablation) | ContextDesc+ · HPSequences | matching performance | No clear effect |
| learn the feature aggregation with concatenation and fully connected layers instead of simple sum and normalization (ablation) | cross-modality feature aggregation · HPatches | performance: 2 points decrease | Hurt |
| aggregate original and doubled patch domain sizes (ablation) | ContextDesc+ · HPSequences | Recall: 67.14 -> 67.74; 76.42 -> 77.51 | Helped |
Closest papers by meaning
- PPFNet: Global Context Aware Local Features for Robust 3D Point Matching (CVPR 2018 · similarity 0.87)
- PUMP: Pyramidal and Uniqueness Matching Priors for Unsupervised Learning of Local Descriptors (CVPR 2022 · similarity 0.87)
- CrossOver: 3D Scene Cross-Modal Alignment (CVPR 2025 · similarity 0.84)
- MutualVPR: A Mutual Learning Framework for Resolving Supervision Inconsistencies via Adaptive Clustering (NeurIPS 2025 · similarity 0.84)
- Context and Geometry Aware Voxel Transformer for Semantic Scene Completion (NeurIPS 2024 · similarity 0.84)
- Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition (AAAI 2025 · similarity 0.84)
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.