Table 1 Detection head module components and their parameter Quantities/FLOPs Contribution.
From: YOLC with dynamic sparse attention for high-speed small target detection in wearable sports images
Sub-component | Parameters (K) | FLOPs (M) |
|---|---|---|
Shared 3 × 3 conv (for classification & regression) | 128.5 | 186.7 |
CA - spatial squeeze | 0.8 | 1.2 |
CA - bottleneck FC layer | 3.6 | 0.9 |
CA - attention map fusion | 0.2 | 2.1 |
Classification branch (1 × 1 conv) | 45.2 | 65.8 |
Regression branch (1 × 1 conv) | 58.3 | 84.5 |
