Abstract
Weakly Supervised Object Localization (WSOL) aims to identify object extents using only image-level labels. A significant hurdle in WSOL is the "classification-localization gap," where models focus on discriminative parts rather than the whole object. We present SD-LocNet, a framework that treats object locations as latent variables and optimizes them via a self-directed strategy. By leveraging a Localization Stability Score (LSS) to identify reliable instances and propagating their features as priors, SD-LocNet achieves state-of-the-art results on PASCAL VOC and MS COCO datasets.
Introduction
The primary challenge in WSOL is that standard classifiers often converge on the most discriminative parts of an object (e.g., a bird's beak) rather than its full spatial extent. This leads to noisy initializations and local minima during training. SD-LocNet addresses this by identifying "easy" images where localization is stable and using them to guide the learning process for more "difficult" or cluttered scenes.
Methodology
SD-LocNet introduces two core innovations to refine noisy object instances:
1. Adaptive Sampling via Localization Stability Score (LSS)
Traditional methods often rely on high-confidence classification scores, which can be misleading. Instead, SD-LocNet measures temporal stability. If a predicted bounding box remains consistent across multiple training epochs, it is assigned a high LSS and considered a "reliable" instance.
2. Reliability Induced Prior Propagation
Features from reliable instances are clustered to form category-specific priors. These priors then regularize the features of "unreliable" instances, forcing them to align with representative category characteristics. This "easy-to-hard" curriculum helps the model discover the full extent of objects in complex images.
Additionally, the network incorporates Multi-context Features using auxiliary convolutional branches to capture varying scales of contextual information, improving background-foreground separation.
Experimental Results
The model was evaluated on three major benchmarks, demonstrating significant improvements over previous state-of-the-art methods like OICR.
- PASCAL VOC 2007: Achieved 70.9% CorLoc (Correct Localization) and 51.3% mAP, surpassing previous methods by approximately 6.6% in CorLoc.
- PASCAL VOC 2012: Reached 69.1% CorLoc, a significant increase over the previous best of 62.1%.
- MS COCO: On this challenging dataset, SD-LocNet improved the baseline localization accuracy from 26.1% to 40.3%.
Conclusion and Future Work
SD-LocNet demonstrates that temporal stability is a more effective signal than raw confidence for weakly supervised learning. By automatically generating a curriculum from "easy" to "hard" images, the model narrows the gap between weakly and fully supervised localization. Future research will focus on improving localization for deformable objects (e.g., humans) and reducing the remaining 15% performance gap compared to fully supervised models.
Summary updated