| --- |
| license: other |
| tags: |
| - heal |
| - horizon |
| - occupancy |
| --- |
| |
| # FlashOcc + HENet + LSS |
|
|
| FlashOcc uses LSS (Lift-Splat-Shoot) view transformation: HENet extracts multi-view camera features; `LSSTransformer` predicts depth distribution and lifts 2D features to 3D voxel space (`depth=45`, `num_points=10`, `bev_size=(40,40,0.625)`); fused via `BevEncoder` (`BiFPN`); `FlashOccDetDecoder`/`BEVOCCHead2D` outputs 18-class 3D occupancy predictions. |
|
|
| --- |
|
|
| ## Deployment Metrics |
|
|
| ### Model Parameters |
|
|
| | Model | Model Input | Backbone | Neck | Model Output | |
| |---|---|---|---|---| |
| | FlashOcc | 6-camera multi-view images `(B,6,3,512,960)` | HENet | FPN + LSSTransformer + BevEncoder | Occupancy grid `(B,C,H,W)` | |
|
|
| ### Accuracy Metrics |
|
|
| | March | Metric | float | calibration | qat | hbm | |
| | --- | --- | --- | --- | --- | --- | |
| | J6M | Occ mIoU (MeanIOU) | 0.3664 | 0.3688 | — | 0.369 | |
|
|
| > Results are based on `march = March.NASH_M` (J6M) configuration; this task has no QAT stage (qat column is `—`). |
| > |
| > 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 | 7.65 | 136.07 | 73.80 | |
| | J6P | 5.62 | 731.18 | 78.10 | |
| | J6B | 30.51 | 33.66 | 82.00 | |
| |
| --- |
| |
| ## Model Overview |
| |
| ### Core Design |
| |
| FlashOcc uses LSS (Lift-Splat-Shoot) view transformation: HENet extracts multi-view camera features; `LSSTransformer` predicts depth distribution and lifts 2D features to 3D voxel space (`depth=45`, `num_points=10`, `bev_size=(40,40,0.625)`); fused via `BevEncoder` (`BiFPN`); `FlashOccDetDecoder`/`BEVOCCHead2D` outputs 18-class 3D occupancy predictions. |
| |
| - **Task type**: BEV 3D occupancy prediction (BEV Occupancy Prediction). |
| - **backbone**: HENet (`type=HENet`, `depth=45`, `num_points=10`), extracts multi-view camera features. |
| - **neck**: `FPN` + `LSSTransformer` (Lift-Splat-Shoot view transformation, `bev_size=(40,40,0.625)`, `grid_size=(128,128)`) + `BevEncoder` (`BiFPN`). |
| - **Occupancy head**: `FlashOccDetDecoder` (`BEVOCCHead2D`, `num_classes=18`, `ignore_index=17`). |
| - **Loss**: `CrossEntropyLoss` (occ seg). |
| - **Model input**: 6-view camera images `(B,6,3,512,960)` (`data_shape=(3,512,960)`). |
| - **Model output**: 18-class 3D occupancy grid `(B,C,H,W)` (`occ3d_seg_class` 18 classes, includes `others`/`ignore_index=17`). |
|
|
| ### Official Repo and Paper |
|
|
| Official repo: https://github.com/Yzichen/FlashOCC |
| Paper: https://arxiv.org/abs/2311.12058 |
|
|
| Note: camera backbone HENet is HEAL-developed; official repo uses a different backbone. |
|
|
| ### Reference |
|
|
| For more J6 chip deployment details, see https://developer.horizon.auto/blog/14099 |
|
|
