Research map / Graph learning / Knowledge and temporal graphs
Infer the Whole from a Glimpse of a Part: Keypoint-Based Knowledge Graph for Vehicle Re-Identification
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 a keypoint-based knowledge graph and transformer-based Invisible Feature Inference Model that infers features of invisible vehicle keypoints from visible ones to build view-invariant vehicle re-identification representations.
Problem
Vehicle re-identification across non-overlapping cameras fails when discriminative parts are occluded or invisible due to viewpoint changes, since global single-image features lack view invariance.
Approach
Model structural relations between 20 vehicle keypoints as a knowledge graph of (head, relation, tail) triplets, and train a transformer to infer invisible keypoint features from visible ones conditioned on those relations, then fuse with baseline global features.
Key results
- On VeRi-776, achieves 83.3% mAP and 97.6% rank-1, the best mAP and rank-1 among compared methods.
- On VERI-Wild Small subset, reaches 86.94 mAP and 96.51 rank-1 versus baseline 85.4 mAP and 95.4 rank-1.
- Ablation on VehicleID shows fusing baseline and transformer-inferred keypoint features (Fbase+Fall (T)) beats single Fbase and single Fall (e.g., 87.97 vs 86.2 rank-1 on Small).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the keypoint inference framework to the baseline | Vehicle re-identification model · VeRi-776 | mAP: 81.7 -> 83.3 | Helped |
| Add the keypoint inference framework to the baseline | Vehicle re-identification model · VehicleID Small | mAP: 90.8 -> 92.5 | Helped |
| Add the keypoint inference framework to the baseline | Vehicle re-identification model · VehicleID Medium | mAP: 87.9 -> 88.8 | Helped |
| Add the keypoint inference framework to the baseline | Vehicle re-identification model · VehicleID Large | mAP: 85.2 -> 86.3 | Helped |
| Add the keypoint inference framework to the baseline | Vehicle re-identification model · VERI-Wild Small | mAP: 85.4 -> 86.94 | Helped |
| Add the keypoint inference framework to the baseline | Vehicle re-identification model · VERI-Wild Medium | mAP: 82.2 -> 83.65 | Helped |
| Add the keypoint inference framework to the baseline | Vehicle re-identification model · VERI-Wild Large | mAP: 76.3 -> 78.41 | Helped |
| Add global baseline features to the inferred keypoint features (ablation) | Vehicle re-identification model · VehicleID Small | rank-1: 50.2 -> 87.97 | Helped |
| Add global baseline features to the inferred keypoint features (ablation) | Vehicle re-identification model · VehicleID Medium | rank-1: 45.9 -> 83.35 | Helped |
| Add global baseline features to the inferred keypoint features (ablation) | Vehicle re-identification model · VehicleID Large | rank-1: 42.2 -> 81.61 | Helped |
| Add inferred keypoint features to visible keypoint features and global features (ablation) | Vehicle re-identification model · VehicleID Small | rank-1: 86.82 -> 87.97 | Helped |
| Replace the transformer with UNet in IFIM (ablation) | Vehicle re-identification model · VehicleID Small | rank-1: 87.51 -> 87.97 | Hurt |
Surprising findings
- Using only the 20 inferred keypoint features (single Fall) performs far worse than the baseline (50.2 rank-1 on VehicleID Small), showing keypoint features alone are insufficient.
Limitations
- Uses about 4.5 GB more GPU memory and ~1.2x inference time than the baseline.
- Depends on a separately fine-tuned keypoint detector (from Khorramshahi et al. 2019) and on 20 annotated keypoints per vehicle.
- Does not beat the best prior method on VeRi-776 rank-5 and VehicleID medium rank-5 (VARID is best on rank-5).
Experimental setup
Datasets: VeRi-776 VehicleID VERI-Wild
Benchmarks: VeRi-776 VehicleID VERI-Wild
Metrics: mAP rank-1 rank-5
Baselines: VARID SRF TransReID SAVER PCRNet DFNet SSBVER GiT
Builds on: IDE TransReID Word2Vec
Closest papers by meaning
- Viewpoint-Aware Attentive Multi-View Inference for Vehicle Re-Identification (CVPR 2018 · similarity 0.88)
- Learning Coarse-to-Fine Structured Feature Embedding for Vehicle Re-Identification (AAAI 2018 · similarity 0.87)
- CLIP-driven View-aware Prompt Learning for Unsupervised Vehicle Re-identification (AAAI 2025 · similarity 0.87)
- Structural Information Guided Multimodal Pre-training for Vehicle-Centric Perception (AAAI 2024 · similarity 0.85)
- Alice Benchmarks: Connecting Real World Re-Identification with the Synthetic (ICLR 2024 · similarity 0.83)
- Unsupervised 3D Structure Inference from Category-Specific Image Collections (CVPR 2024 · similarity 0.82)
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.