| --- |
| license: other |
| tags: |
| - heal |
| - horizon |
| --- |
| |
| # Deformable DETR (ResNet-50) |
|
|
| Deformable DETR replaces DETR's global self-attention with multi-scale deformable attention: each query only samples a small number of points near reference points, enabling faster convergence and lower compute. Four feature levels plus 900 queries provide multi-scale candidates, then the decoder progressively refines boxes layer by layer. |
|
|
| --- |
|
|
| ## Deployment Metrics |
|
|
| ### Model Parameters |
|
|
| | Model | Model Input | Backbone | Neck | Model Output | |
| |---|---|---|---|---| |
| | DeformableDETR | Single image `1x3x800x1332` | ResNet-50 | `ChannelMapperNeck` | Detection boxes `(B,N,cls+reg)` | |
|
|
| ### Accuracy Metrics |
|
|
| | March | Metric | float | calibration | qat | hbm | |
| | --- | --- | --- | --- | --- | --- | |
| | J6M | mAP | 0.4384 | 0.412 | 0.4526 | 0.4529 | |
|
|
| > 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 | 144.84 | 6.92 | 656.00 | |
| | J6P | 78.02 | 28.89 | 672.80 | |
| | J6B | - | - | - | |
| |
| J6B performance is not available for this model. |
| |
| --- |
| |
| ## Model Overview |
| |
| ### Core Design |
| |
| Deformable DETR replaces DETR's global self-attention with multi-scale deformable attention: each query only samples a small number of points near reference points, enabling faster convergence and lower compute. Four feature levels plus 900 queries provide multi-scale candidates, then the decoder progressively refines boxes layer by layer. |
| |
| - **Task type**: 2D object detection (2D Object Detection). |
| - **backbone**: ResNet-50 (`ResNet50`, `include_top=False` removes classification head). |
| - **neck**: `ChannelMapperNeck` (`in_channels=[512,1024,2048]`, `out_channel=256`, 1×1 conv, `extra_convs=1`). |
| - **Position encoding**: `PositionEmbeddingSine` (`num_pos_feats=128`, normalized). |
| - **Transformer**: `DeformableDetrTransformer` (encoder 6 layers + decoder 6 layers, `embed_dim=256`, `num_heads=8`, `feedforward_dim=1024`, `num_feature_levels=4`, `num_queries=900`). |
| - **Post-processing**: `DeformDetrPostProcess` (evaluation selects `select_box_nums_for_evaluation=300` boxes). |
| - **Loss**: `DeformableCriterion`: classification focal loss + L1 bbox + GIoU, `HungarianMatcher` bipartite matching, `aux_loss=True`. |
| - **Key flags**: `with_box_refine=False`, `as_two_stage=False`. |
| - **Model input**: Single image, size `800 × 1332`. |
| - **Model output**: 80-class detection boxes + confidence scores. |
|
|
| ### Official Repo and Paper |
|
|
| Official repo: https://github.com/fundamentalvision/Deformable-DETR |
| Paper: https://arxiv.org/abs/2010.04159 |
|
|
