| --- |
| license: other |
| tags: |
| - heal |
| - horizon |
| - bev |
| --- |
| |
| # BEVFormer (ResNet-50) |
|
|
| BEVFormer 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. |
|
|
| --- |
|
|
| ## Deployment Metrics |
|
|
| ### Model Parameters |
|
|
| | Model | Model Input | Backbone | Neck | Model Output | |
| |---|---|---|---|---| |
| | BevFormer | 6-camera multi-view image sequence `(B,6,3,480,800)` | ResNet-50 | FPN | 3D bounding boxes `(B,N,cls+reg)` | |
|
|
| ### Accuracy Metrics |
|
|
| | March | Metric | float | calibration | qat | hbm | |
| | --- | --- | --- | --- | --- | --- | |
| | J6M | NDS | 0.3739 | 0.3607 | 0.3734 | 0.3669 | |
|
|
| > Results measured with `march = March.NASH_M` (J6M) configuration. |
| > |
| > HEAL version: heal 0.0.2 / hbdk4-compiler 4.11.11 / horizon_plugin_pytorch 3.3.10. |
| |
| ### Performance Metrics |
| |
| > **Performance benchmark**: FPS is measured with single-core 8 threads; latency is single-core single-thread; memory is peak DDR usage. |
| |
| | March | latency (ms) | fps | Memory Usage (MB) | |
| |---|---|---|---| |
| | J6M | 21.80 | 46.62 | 108.20 | |
| | J6P | 14.07 | 277.01 | 108.60 | |
| | J6B | - | - | - | |
| |
| J6B performance is not available for this model. |
| |
| --- |
| |
| ## Model Overview |
| |
| ### Core Design |
| |
| BEVFormer 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. |
| |
| - **Task type**: BEV 3D object detection (BEV 3D Object Detection). |
| - **backbone**: ResNet-50 (`ResNet50`, `include_top=False`, pretrained `num_classes=1000`). |
| - **neck**: FPN (`FPN`, multi-scale feature pyramid, `out_strides=[32]`, `out_channels=[256]`). |
| - **Detection head**: `BEVFormerDetDecoder` (DETR-style decoder + Hungarian matching). |
| - **Loss function**: `BevFormerCriterion` (FocalLoss + L1Loss, matched via BevFormerHungarianAssigner3D). |
| - **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`). |
| - **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`). |
| |
| ### Official Repo and Paper |
| |
| Official repo: https://github.com/fundamentalvision/BevFormer |
| Paper: https://arxiv.org/abs/2203.17270 |
| |
| ### Reference |
| |
| For more J6 chip deployment details, see https://developer.horizon.auto/blog/14101 |
| |
