Fig. 1. Illustration of difficulties in occluded object detection. (a) Loose prediction boxes of heavily overlapped objects; (b) center points of prediction boxes of heavily overlapped objects locate in same feature grid; (c) most regions in occluded object box occupied by foreground object

Fig. 2. Convergence results before and after introducing Tight Loss function. (a) Variance of convergence result of prediction box is relatively larger without Tight Loss function; (b) prediction boxes with different anchor frames as starting points tend to be consistent after introducing Tight Loss function

Fig. 3. Schematic diagrams of high-resolution feature pyramid and insertion position of spatial attention prediction head in network. (a) YOLOv3 network; (b) high-resolution feature pyramid; (c) center points of heavily overlapped objects locate in same grid in original feature pyramid; (d) center points of heavily overlapped objects locate in different grids in high resolution feature pyramid

Fig. 4. Schematic diagram of redundant bounding boxes with high confidence in high-resolution feature pyramid. (a) Target box in original feature pyramid and its confidence prediction; (b) confidence prediction and redundant prediction boxes generated by upsampling mechanism; (c) confidence and prediction boxes filtered by spatial attention mechanism

Fig. 5. Spatial attention module

Fig. 6. Schematic diagrams of spatial attention prediction head and spatial attention residual block. (a) Spatial attention prediction head; (b) spatial attention residual block

Fig. 7. Heat map of target confidence

Fig. 8. Influence of Tight Loss function on model performance. (a)(c) Prediction results after Tight Loss fine-tuning; (b)(d) prediction results without Tight Loss adjustment

Fig. 9. Comparison of comprehensive performance of models. (a)(c) Prediction results generated by improved YOLOv3; (b)(d) prediction results generated by original YOLOv3