[Paper Review] Bridging Saliency Detection to Weakly Supervised Object Detection Based on Self-paced Curriculum Learning

Dingwen Zhang, Deyu Meng+2 more|arXiv (Cornell University)|Mar 3, 2017

Visual Attention and Saliency Detection31 references83 citations

TL;DR

The paper proposes using saliency detection as a prior to weakly supervised object detection and guides learning with self-paced curriculum learning to progressively learn from easy to hard concepts, achieving state-of-the-art results under weak supervision.

ABSTRACT

Weakly-supervised object detection (WOD) is a challenging problems in computer vision. The key problem is to simultaneously infer the exact object locations in the training images and train the object detectors, given only the training images with weak image-level labels. Intuitively, by simulating the selective attention mechanism of human visual system, saliency detection technique can select attractive objects in scenes and thus is a potential way to provide useful priors for WOD. However, the way to adopt saliency detection in WOD is not trivial since the detected saliency region might be possibly highly ambiguous in complex cases. To this end, this paper first comprehensively analyzes the challenges in applying saliency detection to WOD. Then, we make one of the earliest efforts to bridge saliency detection to WOD via the self-paced curriculum learning, which can guide the learning procedure to gradually achieve faithful knowledge of multi-class objects from easy to hard. The experimental results demonstrate that the proposed approach can successfully bridge saliency detection and WOD tasks and achieve the state-of-the-art object detection results under the weak supervision.

Motivation & Objective

• Motivate the challenge of weakly supervised object detection (WOD) where object locations must be inferred from image-level labels.
• Explore how saliency detection can provide useful priors to guide WOD.
• Introduce a self-paced curriculum learning strategy to progressively learn multi-class objects from easy to hard.
• Analyze challenges and limitations in applying saliency to WOD and propose a bridging framework.

Proposed method

• Analyze the role and limitations of saliency detection as a prior for WOD.
• Propose a self-paced curriculum learning framework to guide the WOD learning process from easy to hard concepts.
• Integrate saliency-derived cues into the WOD training pipeline to steer candidate regions.
• Declare the learning strategy to gradually refine object localization under weak supervision.
• Demonstrate that the approach yields improved detection performance under weak supervision.

Experimental results

Research questions

• RQ1 Can saliency detection provide effective priors to improve weakly supervised object detection?
• RQ2 How can self-paced curriculum learning mitigate ambiguity in saliency-guided WOD?
• RQ3 Does bridging saliency with a curriculum learning strategy yield higher-quality object localizations under weak supervision?
• RQ4 What are the main challenges when integrating saliency cues into WOD frameworks?

Key findings

• The approach bridges saliency detection and WOD, enabling improved object detection under weak supervision.
• Self-paced curriculum learning guides the model from easier to harder localization tasks, aiding faithful multi-class learning.
• Experimental results indicate the method achieves state-of-the-art performance for weakly supervised object detection at the time.
• The study highlights the potential of combining human-like selective attention with progressive learning in WOD settings.