Scribble Supervised Annotation Algorithms of Panoptic Segmentation for Autonomous Driving

/in , , /by
Scribble Supervised Annotation Algorithms of Panoptic Segmentation for Autonomous Driving

Shen, Ruobing and Guthier, Thomas and Tang, Bo and Ayed, Ismail Ben

NeurIPS Workshop 2019

Abstract : Large-scale ground truth dataset is of crucial importance for deep learning based segmentation models, but annotating per-pixel masks is extremely time consuming. In this paper, we investigate semi-annotated graph based segmentation algorithms that enforce connectivity. To be more precise, we introduce a class-agnostic heuristic of a discrete Potts model, and a class-aware Integer Linear Programming (ILP) that ensures global optimum. Both algorithms are able to generate panoptic segmentation supervised by scribbles, and can take RGB, or utilize the feature maps from any DCNN, whether trained on the target dataset or not, as input. We present competitive semantic segmentation results on the PASCAL VOC dataset, as well as report panoptic segmentation result on the more challenging Cityscapes dataset. Our algorithms show superior results that makes them suitable for weakly supervised segmentation on new dataset, or interactive semi-automated ground truth generation by human annotators on existing dataset.

You might also like
Model-Based Mean Arterial Pressure Estimation Using Simultaneous Electrocardiogram and Oscillometric Blood Pressure Measurements
An evaluation of over-fit control strategies for multi-objective evolutionary optimization
Unsupervised feature learning for classifying dynamic tactile events using sparse coding
Segmentation of recordings based on partial transcriptions
Automatic Processing of Signature Images: Static Techniques and Methods
Loss factors for learning Boosting ensembles from imbalanced data
Automatic bankcheck processing: information storage and retrieval
Volatility forecasts of the S&P100 by evolutionary programming in a modified time series data mining framework