Hybrid instance segmentation reduces labeling cost with bounding boxes
Hybrid supervised instance segmentation by learning label noise suppression
Computer Vision and Pattern Recognition
Summary
Labeling images pixel by pixel to teach computers to recognize objects is expensive and slow. The authors created a method that mixes cheaper box-shaped labels with some detailed labels to teach the computer without losing much accuracy. They use smart models to fix mistakes in the box labels so the computer learns better. Their approach works well on popular image datasets while cutting down the need for costly detailed labeling.
What this means in practice
- •For computer vision developers: Reduce costly pixel labeling by improving training with mainly bounding box labels for instance segmentation.
- •For autonomous vehicle teams: Improve object recognition training efficiency by combining cheap and accurate labels to speed labeling for scene understanding.
Authors
Linwei Chen, Ying Fu, Shaodi You, Hongzhe Liu
Abstract
To reach top accuracy, current fully supervised instance segmentation methods severely rely on large-scale pixel-wise labeled datasets. They are usually expensive and time-consuming to obtain. Though weakly or semi-supervised methods utilize cheap bounding box labeled, image-level labeled or unlabeled samples to save the labeling cost, their performance is largely sacrificed. To save labeling cost without losing much performance, in this paper, we present a pipeline that can utilize economical bounding box labels and accurate pixel-wise labels in a hybrid way. Specifically, we design two ancillary models to learn label noise suppression and obtain accurate pseudo pixel-wise labels from bounding box labels for training. One is designed to suppress mislabeling between foreground and background, and the other is designed to suppress noise from mislabeling of instances. Moreover, we exploit category-aware spatial attention module, category constraint module, instance constraint module, and self-learning training approach to improve the accuracy of pseudo labels. Experiments on the PASCAL VOC 2012 and the Cityscapes datasets show that our method can achieve competitive performance with much less labeling cost.