Deep Self-Taught Learning for Weakly Supervised Object Localization (1704.05188v2)

Abstract: Most existing weakly supervised localization (WSL) approaches learn detectors by finding positive bounding boxes based on features learned with image-level supervision. However, those features do not contain spatial location related information and usually provide poor-quality positive samples for training a detector. To overcome this issue, we propose a deep self-taught learning approach, which makes the detector learn the object-level features reliable for acquiring tight positive samples and afterwards re-train itself based on them. Consequently, the detector progressively improves its detection ability and localizes more informative positive samples. To implement such self-taught learning, we propose a seed sample acquisition method via image-to-object transferring and dense subgraph discovery to find reliable positive samples for initializing the detector. An online supportive sample harvesting scheme is further proposed to dynamically select the most confident tight positive samples and train the detector in a mutual boosting way. To prevent the detector from being trapped in poor optima due to overfitting, we propose a new relative improvement of predicted CNN scores for guiding the self-taught learning process. Extensive experiments on PASCAL 2007 and 2012 show that our approach outperforms the state-of-the-arts, strongly validating its effectiveness.

• The paper introduces "deep self-taught learning," a novel method improving Weakly Supervised Object Localization (WSL) quality using only image-level annotations to reduce annotation costs.
• This approach employs a threefold strategy including seed sample acquisition linking image annotations to proposals, dense subgraph discovery, and dynamic sample harvesting guided by CNN score improvement.
• Experimental results demonstrate that this method outperforms state-of-the-art WSL techniques on datasets like PASCAL VOC, achieving better average precision and localization accuracy.

Deep Self-Taught Learning for Weakly Supervised Object Localization

In the domain of computer vision, the task of Weakly Supervised Localization (WSL) aims to identify the location of objects within images using only image-level annotations, thus reducing the need for expensive bounding box annotations during the training phase. The paper by Jie et al. investigates this problem and introduces a novel methodology termed "deep self-taught learning" that enhances object localization quality in WSL frameworks by progressively refining detector capabilities.

The conventional approaches in WSL often rely on Multiple Instance Learning (MIL) paradigms to extract promising positive samples, which are inherently limited by the lack of spatial information provided by standard convolutional neural networks (CNNs) trained for classification tasks. This results in detectors that only yield marginal improvements in localization ability. The approach proposed by Jie et al. addresses these limitations through a threefold strategy.

First, they introduce a seed sample acquisition process that links image-level annotations to object proposals utilizing an image-to-object transferring scheme. This scheme ensures the derivation of high-quality seed samples by identifying proposals with significant responses in a multi-label classification network. Then, these proposals undergo a dense subgraph discovery process that selects spatially dense regions as reliable initial samples, thereby reducing the inclusion of spurious or context-laden proposals.

The core innovation is the deep self-taught learning framework, where the training of the detector is reinforced by online supportive sample harvesting. This harvesting is governed by a relative improvement metric of CNN scores, which aids in dynamically selecting the most confident positive samples. Such a strategy deters overfitting by ensuring that only proposals contributing to an improving detector ability are trained further, rather than those simply fitting the initial proposals due to detector bias.

The experimental evidence on PASCAL VOC 2007 and 2012 highlights the efficacy of this approach, often outperforming other state-of-the-art WSL methods across both datasets. Specifically, their method achieves better average precision and localization accuracy, indicative of its robustness and ability to produce high-quality object detectors.

The practical implications of this research are profound; it not only elevates the performance of weakly supervised object detectors but also significantly reduces the annotation burden, making large-scale vision applications more economically feasible. Theoretically, the paper adds to the understanding of self-improvement strategies within neural network training, showcasing a dynamic feedback mechanism where models evolve based on their self-assessed performance metrics.

Future developments in AI could see an expansion of self-taught learning mechanisms in various supervised learning paradigms, where models continuously refine their understanding and adjust their parameters based on adaptive criteria rather than static objectives. Overall, Jie et al.'s contribution paves the way for more refined and cost-effective methods in the field of object localization and detection.