INDIRECT ATTENTION: IA-DETR FOR ONE SHOT OBJECT DETECTION

Bissmella Bahaduri, Hicham Talaoubrid, Fangchen FENG, Zuheng Ming, Anissa Mokraoui

Keywords: One shot object detection, DETR, cross-attention

TL;DR: We introduce a novel type of attention mechanism for three different sequences, and apply it in demanding task of one-shot object detection.

Abstract:

One-shot object detection presents a significant challenge, requiring the identification of objects within a target image using only a single sample image of the object class as query image. Attention-based methodologies have garnered considerable attention in the field of object detection. Specifically, the cross-attention module, as seen in DETR, plays a pivotal role in exploiting the relationships be-
tween object queries and image features. However, in the context of DETR networks for one-shot object detection, the intricate interplay among target image features, query image features, and object queries must be carefully considered.
In this study, we propose a novel module termed ”indirect attention.” We illustrate that relationships among target image features, query image features, and object queries can be effectively captured in a more concise manner compared to
cross-attention. Furthermore, we introduce a pre-training pipeline tailored specifically for one-shot object detection, addressing three primary objectives: identifying objects of interest, class differentiation, and object detection based on a given
query image. Our experimental findings demonstrate that the proposed IA-DETR (Indirect-Attention DETR) significantly outperforms state-of-the-art one-shot object detection methods on both the Pascal VOC and COCO benchmarks.

Primary Area: transfer learning, meta learning, and lifelong learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6371