Enhanced Spatial Feature Learning for Weakly Supervised Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised object detection (WSOD) trains detectors using only class labels, but often results in imprecise localization due to focus on discriminative object parts.
• Existing solutions like custom backbones require extensive pretraining or training from scratch, increasing computational costs and time.
• Pooling layers are critical for spatial feature learning but lack learnable parameters, offering an opportunity for modification without affecting pretrained models.

Purpose of the Study:

• To optimize existing pretrained backbones for WSOD without compromising their original pretraining.
• To introduce a novel method for enhancing spatial feature learning specifically for WSOD challenges.
• To improve the localization accuracy of WSOD detectors by addressing the issue of incomplete object feature learning.

Main Methods:

• Proposed Enhanced Spatial Feature Learning (ESFL) by modifying pooling layers within the backbone.
• ESFL utilizes multiple kernels in pooling layers to address objects of various scales.
• ESFL enhances above-average activations in local neighborhoods to capture less salient object parts.

Main Results:

• ESFL significantly improves the performance of WSOD methods.
• The proposed method achieves state-of-the-art results on benchmark datasets like PASCAL VOC and MS COCO.
• ESFL effectively alleviates the problem of imprecise object localization in WSOD.

Conclusions:

• Modifying pooling layers offers an efficient way to enhance WSOD backbones.
• ESFL provides a practical solution for improving WSOD accuracy and localization.
• The approach demonstrates the potential of targeted spatial feature enhancement for weakly supervised learning tasks.