| --- |
| license: other |
| tags: |
| - heal |
| - horizon |
| --- |
| |
| # FCOS3D (EfficientNet-b0) |
|
|
| FCOS3D extends FCOS anchor-free fully-convolutional detection to 3D: EfficientNet-b0 + BiFPN extract multi-scale features; FCOS3DHead predicts 3D detection boxes (center offset, depth, size, orientation, class) at each feature point; FCOS3DTarget performs 3D target assignment; FCOS3DPostProcess decodes final 3D detection boxes. |
|
|
| --- |
|
|
| ## Deployment Metrics |
|
|
| ### Model Parameters |
|
|
| | Model | Model Input | Backbone | Neck | Model Output | |
| |---|---|---|---|---| |
| | FCOS3D | Single front-view image `(B,3,512,896)` | EfficientNet-b0 | BiFPN | Front-view 3D detection boxes `(B,N,cls+reg)` | |
|
|
| ### Accuracy Metrics |
|
|
| | March | Metric | float | calibration | qat | hbm | |
| | --- | --- | --- | --- | --- | --- | |
| | J6M | NDS | 0.312 | 0.3044 | 0.3099 | 0.3098 | |
| | | mono mAP | 0.2101 | 0.2042 | 0.2067 | 0.207 | |
|
|
| > Results are based on `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 measurement**: FPS is measured with single-core eight-thread; Latency is measured with single-core single-thread; Memory is peak DDR usage. |
| |
| | March | latency (ms) | fps | Memory Usage | |
| |---|---|---|---| |
| | J6M | 1.89 | 652.39 | 18.40 | |
| | J6P | 1.49 | 3310.89 | 19.10 | |
| | J6B | 6.59 | 183.42 | 14.00 | |
| |
| --- |
| |
| ## Model Overview |
| |
| ### Core Design |
| |
| FCOS3D extends FCOS anchor-free fully-convolutional detection to 3D: EfficientNet-b0 + BiFPN extract multi-scale features; FCOS3DHead predicts 3D detection boxes (center offset, depth, size, orientation, class) at each feature point; FCOS3DTarget performs 3D target assignment; FCOS3DPostProcess decodes final 3D detection boxes. |
| |
| - **Task type**: Monocular 3D object detection (Monocular 3D Object Detection). |
| - **backbone**: EfficientNet-b0 (`efficientnet`, `model_type=b0`, `include_top=False`, `activation=relu`, `use_se_block=False`). |
| - **neck**: BiFPN (`BiFPN`, bidirectional feature pyramid, `stack=3`, `out_channels=64`, `num_outs=5`). |
| - **Detection head**: `FCOS3DHead` (fully-convolutional anchor-free 3D detection head). |
| - **Loss**: `FCOS3DLoss` (FocalLoss + SmoothL1Loss + CrossEntropyLoss). |
| - **Model input**: Single front-view image (only CAM_FRONT from 6 cameras), `(B,3,512,896)` (Resize3D `img_scale=(896,512)` + Pad `(512,896)`). |
| - **Model output**: Front-view 3D detection boxes (class + center + size + orientation), per-feature-point prediction `group_reg_dims=(2,1,3,1,2)`, `num_classes=10`, decoded via `FCOS3DPostProcess` + NMS (`max_per_img=100`). |
|
|
| ### Official Repo and Paper |
|
|
| Official repo: https://github.com/open-mmlab/mmdetection3d |
| Paper: https://arxiv.org/abs/2104.10956 |
|
|
| Note: backbone is EfficientNet-b0; official implementation uses a different backbone. |
|
|
| ### Reference |
|
|
| For more J6 chip deployment details, see https://developer.horizon.auto/blog/14096 |
|
|
