OpenExplorer/bevformer_tiny_resnet50_detection

1---2license: other3tags:4  - heal5  - horizon6  - bev7---8 9# BEVFormer (ResNet-50)10 11BEVFormer extracts BEV features from multi-camera sequences via learnable spatiotemporal attention (Temporal Self-Attention + Spatial Cross-Attention): ResNet-50 + FPN extract multi-scale image features, BevFormerViewTransformer projects to BEV, BEVFormerEncoder fuses temporal and spatial information, and BEVFormerDetDecoder decodes 3D bounding boxes.12 13---14 15## Deployment Metrics16 17### Model Parameters18 19| Model | Model Input | Backbone | Neck | Model Output |20|---|---|---|---|---|21| BevFormer | 6-camera multi-view image sequence `(B,6,3,480,800)` | ResNet-50 | FPN | 3D bounding boxes `(B,N,cls+reg)` |22 23### Accuracy Metrics24 25| March | Metric | float | calibration | qat | hbm |26| --- | --- | --- | --- | --- | --- |27| J6M | NDS | 0.3739 | 0.3607 | 0.3734 | 0.3669 |28 29> Results measured with `march = March.NASH_M` (J6M) configuration.30>31> HEAL version: heal 0.0.2 / hbdk4-compiler 4.11.11 / horizon_plugin_pytorch 3.3.10.32 33### Performance Metrics34 35> **Performance benchmark**: FPS is measured with single-core 8 threads; latency is single-core single-thread; memory is peak DDR usage.36 37| March | latency (ms) | fps | Memory Usage (MB) |38|---|---|---|---|39| J6M | 21.80 | 46.62 | 108.20 |40| J6P | 14.07 | 277.01 | 108.60 |41| J6B | - | - | - |42 43J6B performance is not available for this model.44 45---46 47## Model Overview48 49### Core Design50 51BEVFormer extracts BEV features from multi-camera sequences via learnable spatiotemporal attention (Temporal Self-Attention + Spatial Cross-Attention): ResNet-50 + FPN extract multi-scale image features, BevFormerViewTransformer projects to BEV, BEVFormerEncoder fuses temporal and spatial information, and BEVFormerDetDecoder decodes 3D bounding boxes.52 53- **Task type**: BEV 3D object detection (BEV 3D Object Detection).54- **backbone**: ResNet-50 (`ResNet50`, `include_top=False`, pretrained `num_classes=1000`).55- **neck**: FPN (`FPN`, multi-scale feature pyramid, `out_strides=[32]`, `out_channels=[256]`).56- **Detection head**: `BEVFormerDetDecoder` (DETR-style decoder + Hungarian matching).57- **Loss function**: `BevFormerCriterion` (FocalLoss + L1Loss, matched via BevFormerHungarianAssigner3D).58- **Model input**: 6-camera multi-view image sequence, `(B,6,3,480,800)` (original `orig_shape=(3,900,1600)` → resize `(3,450,800)` → pad to `(3,480,800)`, `num_views=6`, training `queue_length=3`, evaluation `queue_length=1`).59- **Model output**: 3D bounding boxes on BEV features (class + center + size + orientation), `num_query=900`, `num_classes=10`, decoded via `BevFormerProcess` (`max_num=300`, `score_threshold=0.3`).60 61### Official Repo and Paper62 63Official repo: https://github.com/fundamentalvision/BevFormer64Paper: https://arxiv.org/abs/2203.1727065 66### Reference67 68For more J6 chip deployment details, see https://developer.horizon.auto/blog/1410169