PointTransformer: Encoding Human Local Features for Small Target Detection

Table 1

Influence of different backbone design on feature encoding.

Backbone design	AP	AP50	AP75	APs	APm	APl
YOLOV5-X	31.4	40.1	33.2	17.3	38.7	53.5
Transformer encoder	31.6	40.4	33.7	18.6	39.6	49.8
Self-attention and cross-attention	34.5	42.1	37.1	24.2	42.1	45.6