--- license: other library_name: pytorch tags: - blindmap - collaborative-perception - cooperative-perception - autonomous-driving - v2x - opv2v - dair-v2x - v2xset - lidar - camera - 3d-object-detection - pytorch - opencood model-index: - name: BlindMap Checkpoints results: - task: type: object-detection name: Cooperative 3D Object Detection dataset: type: OPV2V name: OPV2V metrics: - type: AP@0.3 name: OPV2V Camera AP@0.3 value: 70.03 - type: AP@0.7 name: OPV2V LiDAR AP@0.7 value: 92.03 - type: AP@0.7 name: OPV2V Heterogeneous AP@0.7 value: 79.43 - task: type: object-detection name: Cooperative 3D Object Detection dataset: type: DAIR-V2X-C name: DAIR-V2X-C metrics: - type: AP@0.3 name: DAIR Camera AP@0.3 value: 33.21 - type: AP@0.7 name: DAIR LiDAR AP@0.7 value: 63.97 - type: AP@0.7 name: DAIR Heterogeneous AP@0.7 value: 40.49 - task: type: object-detection name: Cooperative 3D Object Detection dataset: type: V2XSet name: V2XSet metrics: - type: AP@0.3 name: V2XSet Camera AP@0.3 value: 49.65 - type: AP@0.7 name: V2XSet LiDAR AP@0.7 value: 78.71 - type: AP@0.7 name: V2XSet Heterogeneous AP@0.7 value: 68.55 --- # BlindMap Checkpoints This repository contains checkpoints for **BlindMap**, a communication-efficient collaborative perception method for deadline-constrained V2X perception. Paper: > **Deadline-Aware Collaborative Perception via Receiver-Conditioned Utility > Modeling** > Zhenhan Zhu, Yanchao Zhao, Yihang Jiang, Hao Han, and Jie Wu. Code release: - Repository: [AlexZhu2000/BlindMap](https://github.com/AlexZhu2000/BlindMap) - Repository: [NUAA-SmartSensing/BlindMap](https://github.com/NUAA-SmartSensing/BlindMap.git) - Branch: `TMC` - Commit used for this release: `075b2c7120277a619f41bc49a58e3905ffad7aa7` ## Model Layout ```text . ├── README.md ├── LICENSE ├── CHECKPOINT_MANIFEST.yaml ├── SHA256SUMS └── models ├── dair │ ├── camera │ │ ├── config.yaml │ │ └── net_epoch_bestval_at17.pth │ ├── lidar │ │ ├── blindmap_quality_result.txt │ │ ├── config.yaml │ │ └── net_epoch_bestval_at21.pth │ └── heterogeneous │ ├── config.yaml │ └── net_epoch_bestval_at35.pth ├── opv2v │ ├── camera │ │ ├── config.yaml │ │ └── net_epoch_bestval_at17.pth │ └── lidar_heterogeneous │ ├── blindmap_quality_result.txt │ ├── config.yaml │ └── net_epoch_bestval_at37.pth └── v2xset ├── camera │ ├── config.yaml │ └── net_epoch_bestval_at19.pth ├── lidar │ ├── blindmap_quality_result.txt │ ├── config.yaml │ └── net_epoch_bestval_at19.pth └── heterogeneous ├── config.yaml └── net_epoch_bestval_at23.pth ``` The DAIR-V2X-C and V2XSet releases use separate checkpoints for camera-only, LiDAR-only, and heterogeneous settings. OPV2V is organized differently: the LiDAR-only profile is evaluated from the LiDAR branch inside the heterogeneous m1m2 checkpoint, following the HEAL-style heterogeneous setup. Therefore OPV2V provides `camera` and `lidar_heterogeneous` directories rather than a separate `lidar` directory. ## Checkpoints | Profile | Model directory | Checkpoint | SHA-256 | |---|---|---|---| | DAIR camera-only | `models/dair/camera` | `net_epoch_bestval_at17.pth` | `b95722f3b58a13a979315a84133da971415cdc9b6a94bb555b013bd834576d81` | | DAIR LiDAR-only | `models/dair/lidar` | `net_epoch_bestval_at21.pth` | `99302097362f02dfb73f251a4bdf7d0c4e2d7ea9ebbc51422eb322d3f27e4429` | | DAIR heterogeneous | `models/dair/heterogeneous` | `net_epoch_bestval_at35.pth` | `10511d4a9419f42051f01be2b9e6313b6718b2843f74795b5d9ceac7b200f7f5` | | OPV2V camera-only | `models/opv2v/camera` | `net_epoch_bestval_at17.pth` | `33ddf54fe56d82d2719876a362b4838a4f21a14a068e9acd9c16066c41800fb3` | | OPV2V LiDAR-only profile | `models/opv2v/lidar_heterogeneous` | `net_epoch_bestval_at37.pth` | `9e2570b64335c99241dc47fc3aa6cc9f97180a78b3ff5b54330cbf5a2072b7f3` | | OPV2V heterogeneous | `models/opv2v/lidar_heterogeneous` | `net_epoch_bestval_at37.pth` | `9e2570b64335c99241dc47fc3aa6cc9f97180a78b3ff5b54330cbf5a2072b7f3` | | V2XSet camera-only | `models/v2xset/camera` | `net_epoch_bestval_at19.pth` | `0dccab286798a380a78257ce7559e7e7978295a67e42d386e558f9f397e4edca` | | V2XSet LiDAR-only | `models/v2xset/lidar` | `net_epoch_bestval_at19.pth` | `0a0553ae5e457071f39eb13dedf5f9627ad7d5da6a73898ecf0b9dbb6adcc230` | | V2XSet heterogeneous | `models/v2xset/heterogeneous` | `net_epoch_bestval_at23.pth` | `60a73178b7178a18ea59e7beb8f0b4e7c1d115df3d7a97b9a4ebc511dc6d7bb9` | The table lists SHA-256 checksums for checkpoint weights. The DAIR LiDAR and V2XSet LiDAR values are taken from the corresponding Hugging Face LFS metadata. ## Installation Install the BlindMap codebase first. The checkpoints are designed for the BlindMap/OpenCOOD-style `--model_dir` loader, where each checkpoint directory contains both `config.yaml` and `net_epoch_bestval_at*.pth`. ```bash git clone --branch TMC https://github.com/AlexZhu2000/BlindMap.git cd BlindMap conda create -n blindmap python=3.8 conda activate blindmap pip install -r requirements.txt python setup.py develop python opencood/utils/setup.py build_ext --inplace ``` Use the same CUDA, PyTorch, and `spconv` versions described in the BlindMap repository. The checkpoints were produced in the original BlindMap/OpenCOOD environment and may not be compatible with arbitrary `spconv` versions. ## Data The datasets are not included. Download OPV2V, DAIR-V2X-C, and V2XSet from their official providers and update the dataset paths in the selected `config.yaml` before inference. Typical keys to update: ```yaml root_dir: /path/to/dataset/train_or_train_json validate_dir: /path/to/dataset/validate_or_val_json test_dir: /path/to/dataset/test_or_val_json data_dir: /path/to/dataset/root ``` Not every config uses all of these keys. ## Inference ### Camera-Only ```bash CUDA_VISIBLE_DEVICES=0 python opencood/tools/inference.py \ --model_dir /path/to/BlindMap/models/opv2v/camera \ --fusion_method intermediate \ --modal 1 \ --comm_volume_MB 1 \ --range 102.4,102.4 ``` Use `models/dair/camera` with `--range 102.4,51.2` for DAIR camera-only, and use `models/v2xset/camera` with `--range 102.4,102.4` for V2XSet camera-only. ### LiDAR-Only For OPV2V, use the LiDAR branch of the heterogeneous checkpoint: ```bash CUDA_VISIBLE_DEVICES=0 python opencood/tools/inference.py \ --model_dir /path/to/BlindMap/models/opv2v/lidar_heterogeneous \ --fusion_method intermediate \ --modal 0 \ --comm_volume_MB 1 \ --range 102.4,102.4 ``` For DAIR-V2X-C and V2XSet, use their separate `lidar` checkpoint directories: `models/dair/lidar` with `--range 102.4,51.2`, and `models/v2xset/lidar` with `--range 102.4,102.4`. ### Heterogeneous ```bash CUDA_VISIBLE_DEVICES=0 python opencood/tools/inference.py \ --model_dir /path/to/BlindMap/models/opv2v/lidar_heterogeneous \ --fusion_method intermediate \ --modal 4 \ --comm_volume_MB 1 \ --range 102.4,102.4 ``` Use `models/dair/heterogeneous` with `--range 102.4,51.2` for DAIR heterogeneous, and use `models/v2xset/heterogeneous` with `--range 102.4,102.4` for V2XSet heterogeneous. ## Intended Use These checkpoints are intended for academic research on collaborative perception, communication-efficient feature sharing, and reproducibility studies on OPV2V, DAIR-V2X-C, and V2XSet. They are not intended for deployment in autonomous vehicles or safety-critical systems. ## Limitations - Results depend on the exact BlindMap code revision, dataset split, sensing range, communication-budget accounting, and environment. - OPV2V LiDAR-only results use the LiDAR branch of the OPV2V m1m2 checkpoint; DAIR-V2X-C and V2XSet LiDAR-only results use separate LiDAR checkpoints. - OPV2V and V2XSet are simulated datasets. DAIR-V2X-C is real-world but still does not cover all deployment conditions. - The model card provides representative metadata for provenance. Users should rerun inference and report their own reproduced metrics. ## License The included `LICENSE` file follows the current BlindMap source license. It is an academic research license with redistribution restrictions. If you intend to redistribute, modify, or use these checkpoints outside academic research, obtain the required permission from the rights holders first. Dataset licenses apply separately. ## Citation ```bibtex @article{zhu2026deadline, title = {Deadline-Aware Collaborative Perception via Receiver-Conditioned Utility Modeling}, author = {Zhu, Zhenhan and Zhao, Yanchao and Jiang, Yihang and Han, Hao and Wu, Jie}, journal = {IEEE Transactions on Mobile Computing}, year = {2026}, note = {Manuscript under review} } ```