ESD-HOI: Efficient and Scale-aware Detection for End-to-End Human Object Interactions
Meijun Wu, Shiwei Ma · 2024
Recently, transformer-based Human-Object Interaction (HOI) detection models have shown impressive accuracy on two public benchmarks, HICO-Det and V-COCO. However, the prevailing approaches are single-scale detectors, focusing on utilizing prior language knowledge and causal reasoning rather than the transformer. In this paper, we propose a novel architecture, ESD-HOI, designed to enhance HOI detection performance on small targets and accelerate convergence. ESD-HOI consists of an encoder and cascaded decoders. The model utilizes four-dimensional vectors as queries to improve query initialization. Furthermore, noise is incorporated during the comparative learning of the decoder query to enhance the distinction between the target and the background. This model achieved an outstanding 54.4 mAP in just 15 epochs on the V-COCO dataset, demonstrating exceptional capabilities in small target detection.