| --- |
| license: other |
| tags: |
| - heal |
| - horizon |
| --- |
| |
| # FCOS (EfficientNet-b3) |
|
|
| FCOS eliminates anchors and predicts classification scores, centerness, and distances to box edges per pixel; `BiFPN` performs bidirectional feature fusion across 5 scales to improve small-object detection; training jointly supervises classification, centerness, and regression losses. |
|
|
| --- |
|
|
| ## Deployment Metrics |
|
|
| ### Model Parameters |
|
|
| | Model | Model Input | Backbone | Neck | Model Output | |
| |---|---|---|---|---| |
| | FCOS | Single image `1x3x896x896` | EfficientNet-b3 | `BiFPN` | Detection boxes `(B,N,cls+reg)` | |
|
|
| ### Accuracy Metrics |
|
|
| | March | Metric | float | calibration | qat | hbm | |
| | --- | --- | --- | --- | --- | --- | |
| | J6M | mAP | 0.4802 | 0.4634 | 0.4773 | 0.4772 | |
|
|
| > 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 | 3.90 | 276.48 | 30.90 | |
| | J6P | 2.98 | 1439.97 | 35.10 | |
| | J6B | - | - | - | |
| |
| J6B performance is not available for this model. |
| |
| --- |
| |
| ## Model Overview |
| |
| ### Core Design |
| |
| FCOS eliminates anchors and predicts classification scores, centerness, and distances to box edges per pixel; `BiFPN` performs bidirectional feature fusion across 5 scales to improve small-object detection; training jointly supervises classification, centerness, and regression losses. |
| |
| - **Task type**: 2D object detection (2D Object Detection). |
| - **backbone**: EfficientNet-b3 (`efficientnet`, `model_type=b3`, `include_top=False` removes classification head, `use_se_block=False`, ReLU activation). |
| - **neck**: `BiFPN` (`out_channels=120`, `stack=6`, input stride `[2,4,8,16,32]` → output stride `[8,16,32,64,128]`, `bifpn_sum` fusion). |
| - **Detection head**: `FCOSHead` (`feat_channels=120`, `stacked_convs=4`, 5 feature layers stride `[8,16,32,64,128]`, `dequant_output=True`). |
| - **Target assignment**: `DynamicFcosTarget` (`topK=10`, classification `FocalLoss`, regression `GIoULoss`). |
| - **Post-processing**: `FCOSDecoder` (`score_thr=0.05`, `nms_pre=1000`, NMS `iou_threshold=0.6`, `max_per_img=100`, uses centerness). |
| - **Loss**: Classification `FocalLoss` (`alpha=0.25, gamma=2.0`) + centerness `CrossEntropyLoss` (sigmoid) + regression `GIoULoss`. |
| - **Model input**: Single image, size `896 × 896`. |
| - **Model output**: 80-class detection boxes + confidence scores. |
|
|
| ### Official Repo and Paper |
|
|
| Official repo: https://github.com/tianzhi0549/FCOS |
| Paper: https://arxiv.org/abs/1904.01355 |
|
|
| Note: backbone is EfficientNet-b3; official repo uses a different backbone. |
|
|
