Geometry-aware vehicle re-identification on the VeRi dataset. This project extends a standard ResNet vehicle ReID pipeline with local part descriptors, a pseudo-depth/visibility attention branch, BNNeck retrieval features, camera auxiliary supervision, and optional flip-consistency training.
- Implemented a geometry-aware ReID model registered as
geovehicle_resnet18,geovehicle_resnet34, andgeovehicle_resnet50. - Trained with identity classification and hard triplet loss for retrieval-oriented embeddings.
- Added local horizontal part descriptors to retain vehicle identity cues such as lights, windows, roof shape, wheels, and color regions.
- Added a pseudo-depth/visibility branch to improve viewpoint-sensitive retrieval.
- Evaluated using CMC and mAP on VeRi query/gallery retrieval.
| Model | Dataset | mAP | Rank-1 | Rank-5 | Rank-10 |
|---|---|---|---|---|---|
| GeoVehicle ResNet-18 | VeRi | 60.2% | 84.7% | 94.3% | 96.4% |
The verified run used geovehicle_resnet18, ImageNet pretraining, RandomIdentitySampler, 4 instances per identity, AMSGrad, batch size 64, camera auxiliary loss, flip consistency, L2-normalized features, and flip-test evaluation.
main.py: training and evaluation entry point.args.py: command-line options.src/models/geovehicle.py: geometry-aware ReID architecture.src/losses/: cross-entropy and hard triplet losses.src/eval_metrics.py: CMC and mAP retrieval metrics.train_geovehicle_resnet18.sh: main reproduced training command.run_geovehicle_ablation.sh: ablation runner for model components.docs/geovehicle_reid.md: method notes and ablation guidance.results/verified_results.md: compact result record.
The VeRi dataset and trained checkpoints are not included. Set DATA_ROOT to the dataset root before running scripts.
DATA_ROOT=/path/to/veri-dataset-root bash train_geovehicle_resnet18.shThis project does not claim metric 3D reconstruction from VeRi. The geometry branch is a pseudo-3D/visibility representation designed to improve viewpoint-aware vehicle retrieval when calibrated 3D geometry is unavailable.