Attend Refine Repeat: Active Box Proposal Generation via In-Out Localization

Abstract: The problem of computing category agnostic bounding box proposals is utilized as a core component in many computer vision tasks and thus has lately attracted a lot of attention. In this work we propose a new approach to tackle this problem that is based on an active strategy for generating box proposals that starts from a set of seed boxes, which are uniformly distributed on the image, and then progressively moves its attention on the promising image areas where it is more likely to discover well localized bounding box proposals. We call our approach AttractioNet and a core component of it is a CNN-based category agnostic object location refinement module that is capable of yielding accurate and robust bounding box predictions regardless of the object category. We extensively evaluate our AttractioNet approach on several image datasets (i.e. COCO, PASCAL, ImageNet detection and NYU-Depth V2 datasets) reporting on all of them state-of-the-art results that surpass the previous work in the field by a significant margin and also providing strong empirical evidence that our approach is capable to generalize to unseen categories. Furthermore, we evaluate our AttractioNet proposals in the context of the object detection task using a VGG16-Net based detector and the achieved detection performance on COCO manages to significantly surpass all other VGG16-Net based detectors while even being competitive with a heavily tuned ResNet-101 based detector. Code as well as box proposals computed for several datasets are available at:: https://github.com/gidariss/AttractioNet.

Attend Refine Repeat: Active Box Proposal Generation via In-Out Localization

The paper "Attend Refine Repeat: Active Box Proposal Generation via In-Out Localization" by Spyros Gidaris and Nikos Komodakis addresses the problem of category agnostic bounding box proposal generation, a critical component in object detection tasks within computer vision. The authors introduce AttractioNet, a novel approach that leverages active box proposal generation starting from seed boxes uniformly distributed across an image. The approach progressively refines these boxes by focusing on promising areas likely to contain objects.

Summary of the Approach

The proposed AttractioNet framework hinges on two pivotal innovations:

1. Category Agnostic Object Location Refinement: This component leverages advances from LocNet, enhancing the ability to localize objects accurately regardless of their categories. The refinement module replaces traditional bounding box regression with dense membership probability assignments per row and column, leading to more precise box predictions.
2. Active Box Proposal Generation Strategy: Known as the Attend Refine Repeat algorithm, this strategy iteratively refines the placement of bounding boxes starting from seed boxes. Instead of processing a static set, boxes are dynamically adjusted to better enclose objects, enhancing both their objectness scores and localization.

Experimental Results

The researchers evaluate AttractioNet across several datasets including COCO, PASCAL, ImageNet detection, and NYU-Depth V2, consistently achieving state-of-the-art average recall (AR) rates. On the COCO dataset, for instance, AttractioNet's AR@1000 reaches 66.2%, significantly surpassing prior methods such as SharpMask and EdgeBoxes. Its ability to generalize to unseen categories is validated on datasets like ImageNet detection and NYU-Depth where it maintains high accuracy, outperforming competitive techniques.

Findings demonstrate that AttractioNet not only improves object localization but also boosts object detection systems' performance. Utilizing AttractioNet proposals with a VGG16-Net-based detector, the approach delivers strong detection results on COCO test-dev, rivaling ResNet-101-based configurations.

Implications and Future Directions

The proposed method has far-reaching implications for computer vision applications that require efficient and accurate bounding box generation. The innovations introduced may enhance not only object detection but also visual tasks such as semantic role labeling, caption generation, and visual question answering, which benefit from robust box proposals.

Future work could explore integrating AttractioNet with advanced architectures like Deep Residual Networks to potentially improve detection accuracy further. Moreover, strategies to optimize computational efficiency and runtime, such as network architecture tuning, could make AttractioNet even more practical for deployment in real-world scenarios.

Overall, the paper's contributions set a notable benchmark in bounding box proposal generation, paving the way for refined methodologies in object detection and recognizing the diverse objects present in complex visual scenes.