C-MIL: Continuation Multiple Instance Learning for Weakly Supervised Object Detection (1904.05647v1)

Abstract: Weakly supervised object detection (WSOD) is a challenging task when provided with image category supervision but required to simultaneously learn object locations and object detectors. Many WSOD approaches adopt multiple instance learning (MIL) and have non-convex loss functions which are prone to get stuck into local minima (falsely localize object parts) while missing full object extent during training. In this paper, we introduce a continuation optimization method into MIL and thereby creating continuation multiple instance learning (C-MIL), with the intention of alleviating the non-convexity problem in a systematic way. We partition instances into spatially related and class related subsets, and approximate the original loss function with a series of smoothed loss functions defined within the subsets. Optimizing smoothed loss functions prevents the training procedure falling prematurely into local minima and facilitates the discovery of Stable Semantic Extremal Regions (SSERs) which indicate full object extent. On the PASCAL VOC 2007 and 2012 datasets, C-MIL improves the state-of-the-art of weakly supervised object detection and weakly supervised object localization with large margins.

Continuation Multiple Instance Learning for Weakly Supervised Object Detection

The paper "C-MIL: Continuation Multiple Instance Learning for Weakly Supervised Object Detection" addresses the intricate challenge of weakly supervised object detection (WSOD), where object locations and detectors are derived from image-level category labels rather than precise bounding box annotations. The authors introduce a novel approach, Continuation Multiple Instance Learning (C-MIL), which incorporates continuation optimization into the conventional Multiple Instance Learning (MIL) framework to manage the non-convexity problem that frequently causes localization errors by getting trapped in local minima.

Core Contribution and Methodology

C-MIL's primary innovation lies in the use of continuation methods, transforming a non-convex optimization problem into a series of smoother sub-problems. This technique systematically addresses the non-convex nature of loss functions in MIL, which are susceptible to local minima.

The proposed method achieves this by:

1. Instance Subset Partitioning: Instances within each image are divided into subsets based on spatial and class-related features. This restructuring aids in smoothing the original loss function into a sequence of smoothed versions, thus reducing the chances of prematurely converging to local minima.
2. Gradual Optimization: By utilizing a continuation parameter, the loss function gradually shifts from a smooth, convex form to the original non-convex form, thereby enabling a more effective global optimization approach.
3. Stable Semantic Extremal Regions (SSERs): As C-MIL progresses through the continuation parameters, it discovers regions that reflect the full extent of the object, unlike traditional MIL that often yields incomplete object localizations.

Results and Implications

The approach was evaluated on the PASCAL VOC 2007 and 2012 datasets, where C-MIL demonstrated substantial improvements over state-of-the-art WSOD methods. The notable increase in mAP (mean Average Precision) values by large margins underscores the effectiveness of alleviating non-convexity issues.

• Quantitative Performance: C-MIL exceeded existing methods by up to 3.2% in detection mAP on PASCAL VOC 2007, while similar improvements were observed on VOC 2012, showcasing the method's robustness across datasets.
• Theoretical Implication: This method highlights that systematic optimization approaches, like continuation methods, could be pivotal in handling non-convex loss landscapes that are pervasive in weak supervision tasks.

Future Directions

The introduction of continuation optimization in WSOD opens multiple avenues for future research. Investigating its application to other areas of weak supervision, or extending it to fully supervised tasks to enhance the learning of more accurate models, could be promising directions. Moreover, exploring adaptive continuation schemes that tailor the optimization process dynamically based on dataset characteristics might yield further performance enhancements.

In summary, C-MIL provides a structured and theoretically sound approach to tackle one of the fundamental challenges in weakly supervised object detection. Its success implies potential for broader applications within AI, particularly in domains constrained by limited annotation resources.