| --- |
| license: other |
| tags: |
| - heal |
| - horizon |
| --- |
| |
| # Detr3D (EfficientNet-b3) |
|
|
| Detr3D brings the DETR paradigm to 3D detection: EfficientNet-b3 + BiFPN extract multi-scale image features; view transformation projects 2D features to 3D space; Detr3dTransformer iteratively samples multi-view features with learnable queries and predicts 3D detection boxes; Detr3dTarget performs Hungarian matching during training. |
|
|
| --- |
|
|
| ## Deployment Metrics |
|
|
| ### Model Parameters |
|
|
| | Model | Model Input | Backbone | Neck | Model Output | |
| |---|---|---|---|---| |
| | Detr3D | 6-camera multi-view images `(B,6,3,512,1408)` | EfficientNet-b3 | BiFPN | 3D detection boxes `(B,N,cls+reg)` | |
|
|
| ### Accuracy Metrics |
|
|
| | March | Metric | float | calibration | qat | hbm | |
| | --- | --- | --- | --- | --- | --- | |
| | J6M | NDS | 0.3357 | 0.3299 | 0.338 | 0.337 | |
| | | mAP | 0.2694 | 0.2618 | 0.2688 | 0.2683 | |
|
|
| > 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 | 21.88 | 46.43 | 97.60 | |
| | J6P | 15.22 | 253.24 | 94.60 | |
| | J6B | - | - | - | |
| |
| J6B performance is not available for this model. |
| |
| --- |
| |
| ## Model Overview |
| |
| ### Core Design |
| |
| Detr3D brings the DETR paradigm to 3D detection: EfficientNet-b3 + BiFPN extract multi-scale image features; view transformation projects 2D features to 3D space; Detr3dTransformer iteratively samples multi-view features with learnable queries and predicts 3D detection boxes; Detr3dTarget performs Hungarian matching during training. |
| |
| - **Task type**: BEV 3D object detection (BEV 3D Object Detection). |
| - **backbone**: EfficientNet-b3 (`efficientnet`, `model_type=b3`, `include_top=False`, `activation=relu`, `use_se_block=False`). |
| - **neck**: BiFPN (`BiFPN`, bidirectional feature pyramid, `stack=3`, `out_channels=256`, `num_outs=5`). |
| - **Detection head**: `Detr3dHead` + `Detr3dTransformer` + `Detr3dDecoder` (DETR-style 3D decoder). |
| - **Loss**: FocalLoss (cls) + L1Loss (bbox), via Detr3dTarget Hungarian matching. |
| - **Model input**: 6-camera multi-view images, `(B,6,3,512,1408)` (original `orig_shape=(3,900,1600)` → resize `(3,792,1408)` → crop `data_shape=(3,512,1408)`, `num_views=6`). |
| - **Model output**: 3D detection boxes (class + center + size + orientation), `num_query=900`, `num_classes=10`, decoded via `Detr3dPostProcess` (`max_num=300`). |
|
|
| ### Official Repo and Paper |
|
|
| Official repo: https://github.com/WangYueFt/detr3d |
| Paper: https://arxiv.org/abs/2110.06922 |
|
|
| Note: backbone is EfficientNet-b3; official repo uses a different backbone. |
|
|
