Universal Bounding Box Regression and Its Applications

[howardyclo]

Metadata

• Authors: Seungkwan Lee, Suha Kwak, and Minsu Cho
• Organization: Dept. of Computer Science and Engineering, POSTECH, Korea.
• Conference: ACCV 2018
• Paper: https://arxiv.org/pdf/1904.06805.pdf

[howardyclo]

TL;DR

• Propose a class-agnostic (transferable to unseen classes) and anchor-free box regressor, Universal Bounding-Box Regressor (UBBR).
• UBBR takes an image and any arbitrary bounding boxes, and refines the boxes so that they enclose their nearest objects tightly, regardless of their classes.
• UBBR can also be used to generate object box proposals. Given boxes uniformly and densely sampled from image space, UBBR transforms them to approximate the boxes of their nearest objects.
• Perform well on unseen classes on weakly supervised object detection, object proposals, and object discovery.

Problem

• In weakly supervised object detection, box annotations for target object classes are not given, only image-level class labels. As as result, object bounding boxes tend to be badly localized due to the limited supervision and only discriminative part is covered.

Method

• Data: Any existing object detection dataset with image and its ground-truth boxes.
• Train: Randomly generate roughly localized and diverse boxes: Apply random transform to ground-truth boxes and its IoU should be greater than a threshold.
• Loss: Use IoU loss between ground-truth box and random box, instead of conventional L2 and smooth L1 losses (IoU loss is better). Add tiny constant to loss to enable numerical stability
• b: input box; UBBR(b): offsets predicted by UBBR; f(.): refine b with predicted offsets.

Experiment

• Dataset: Tested on VOC; trained on COCO-17 (remove VOC classes: 80-20=60).
• Use ImageNet-pretrained ResNet101. (Doesn't ImageNet containing "unseen" objects make experiments invalid?)
• Baselines:
  • Weakly supervised object detection:
    • OICR Iteratively refines localization.
  • Object proposals: Out-performed previous 13 methods. Note that unlike many other methods (except SelectiveSearch [18]), UBBR does not use any images from VOC object classes for training.
  • Object discovery:
    • Unsupervised object discovery and localization in the wild: Part-based matching with bottom-up region proposals.
    • Image co-localization by mimicking a good detectors confidence score distribution.

Limitation

• UBBR applied on OICR outputs may not be able to localize nested objects since UBBR is class-agnostic.