File size: 2,823 Bytes
186d1c5 33d5d6a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 | ---
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
|