OpenExplorer/bevformer_henet_camera_multitask

BevFormer+HENet Multitask

The BevFormer+HENet multitask model uses HENet as the camera backbone to extract multi-view features, then BevFormer ViewTransformer and temporal BEV Encoder to transform them into bird's-eye-view features. BEV features are fed to both a detection head (3D object detection) and an occupancy head (FlashOcc-style semantic occupancy prediction), enabling joint det+occ training.

Deployment Metrics

Model Parameters

Accuracy Metrics

Results measured with march = March.NASH_M (J6M) configuration. HEAL version: heal 0.0.2 / hbdk4-compiler 4.11.11 / horizonpluginpytorch 3.3.10.

Model Overview

Core Design

The BevFormer+HENet multitask model uses HENet as the camera backbone to extract multi-view features, then BevFormer ViewTransformer and temporal BEV Encoder to transform them into bird's-eye-view features. BEV features are fed to both a detection head (3D object detection) and an occupancy head (FlashOcc-style semantic occupancy prediction), enabling joint det+occ training.

• —Task type: Multitask fusion (3D object detection + occupancy grid prediction).
• —backbone: HENet (camera feature extraction), extracting multi-view features.
• —neck: FPN.
• —Detection head: BevFormer detection head, outputting 10-class 3D bounding boxes + velocity (num_classes=10).
• —Occupancy head: Semantic occupancy prediction, 18 classes (num_classes_occ=18).
• —BEV range: bev_size=(51.2, 51.2, 0.8), bev_size_occ=(40, 40).
• —Model input: 6-camera image sequence (B,6,3,512,960) (queue_length=1).
• —Model output: det 3D bounding boxes + occ occupancy grid semantics (B,C,H,W).

Official Repo and Paper

Official repo: https://github.com/fundamentalvision/BevFormer Paper: https://arxiv.org/abs/2203.17270

Note: The camera backbone HENet is a HEAL in-house implementation; the official repo uses a different backbone.