Layer-Wise Query Selection to Eliminate Redundant Queries in DETR
Seok-Jin Hong, Chan-Young Choi, Sang-Woong Lee · Applied Sciences · 2025
Recent advancements in the detection Transformer model have demonstrated remarkable accuracy in real-time object detection using an end-to-end approach. DETR leverages the concept of object queries, which act as “questions” to determine the presence and location of objects. However, the excessive number of object queries significantly increases computational complexity, leading to higher training and inference times, greater memory consumption, and increased costs. To address this issue, this paper introduces the selected query detection Transformer, a novel approach that optimizes the selection of object queries in the decoder. By progressively filtering and reducing unnecessary queries, the selected query detection Transformer maintains model performance while significantly reducing computational overhead, resulting in faster training and inference times.