关键词: 人物交互检测, 目标检测, 相对位置编码, 频域特征, 离散余弦变换

Abstract: The task of human-object interaction detection aims to identify all interactions between humans and objects in an image. Most existing approaches employ an encoder-decoder framework for end-to-end training, which heavily relies on absolute positional encoding and performs suboptimally in complex multi-object interaction scenarios. To address the limitations of capturing relative spatial relationships between humans and objects due to reliance on absolute positional encoding, as well as the insufficient integration of local and global information, a novel human-object interaction detection model was proposed by combining cross dimensional interaction feature extraction with frequency-domain feature fusion. The model improves the conventional Transformer encoder by incorporating an additional positional encoding scheme. Through the fusion of relative and absolute positional encodings, the model is enabled to capture the spatial relationships between humans and objects. Furthermore, a new feature extraction module was introduced to enhance image information representation by capturing interactions across channel, spatial, and feature dimensions, while discrete cosine transform was applied to extract frequency-domain features for richer local and global information representation. Finally, the Wise-IoU loss function was adopted to improve detection accuracy and class discrimination capability, allowing the model to flexibly handle targets of various categories. Experiments conducted on two public datasets, HICO-DET and V-COCO, show that the proposed model achieves an improvement of 0.95 percentage points in mAP on the full set of HICO-DET and 0.9 percentage points in AP on scenario 1 of the V-COCO dataset, compared to the GEN-VLKT model.

Key words: human-object interaction detection, object detection, relative position encoding, Frequency domain characteristics, discrete cosine transform