Proposals Generation for Weakly Supervised Object Detection in Artwork Images.

Area of Science:

• Computer Vision
• Machine Learning
• Digital Art History

Background:

• Accurate object detection typically requires precise bounding box annotations, which are scarce for non-natural image datasets like artworks.
• Existing state-of-the-art end-to-end object detection models may perform poorly on datasets that differ significantly from their pre-training data.

Purpose of the Study:

• To develop a novel two-stage Weakly Supervised Object Detection (WSOD) approach for accurate bounding box generation in non-natural image datasets.
• To overcome the limitations of manual annotation scarcity in specialized domains such as art datasets.
• To improve object detection performance on artworks datasets compared to existing WSOD methods.

Main Methods:

• Proposed a two-stage WSOD method leveraging Class Activation Maps (CAMs) to generate pseudo-ground truth bounding boxes.
• Utilized existing image classification knowledge to guide the generation of annotations.
• Trained a Faster R-CNN object detector using the automatically generated pseudo-ground truth bounding boxes.

Main Results:

• Generated bounding boxes from CAMs effectively compensated for the lack of manual ground truth.
• The proposed method surpassed state-of-the-art end-to-end WSOD approaches on the ArtDL 2.0 dataset (≈41.5% mAP) and IconArt dataset (≈17% mAP).
• Demonstrated the viability of pseudo-ground truth for training robust object detectors on specialized datasets.

Conclusions:

• The presented WSOD approach offers a practical solution for accurate object localization in non-natural image datasets lacking manual annotations.
• This method advances computer-aided analysis of artworks and facilitates future research in areas like automatic image captioning for digital archives.