Two Tricks to Improve Unsupervised Segmentation Learning

We present two practical improvement techniques for unsupervised segmentationlearning. These techniques address limitations in the resolution and accuracyof predicted segmentation maps of recent state-of-the-art methods. Firstly, weleverage image post-processing techniques such as guided filtering to refinethe output masks, improving accuracy while avoiding substantial computationalcosts. Secondly, we introduce a multi-scale consistency criterion, based on ateacher-student training scheme. This criterion matches segmentation maskspredicted from regions of the input image extracted at different resolutions toeach other. Experimental results on several benchmarks used in unsupervisedsegmentation learning demonstrate the effectiveness of our proposed techniques.