Object Counting and Instance Segmentation with Image-level Supervision

Common object counting in a natural scene is a challenging problem incomputer vision with numerous real-world applications. Existing image-levelsupervised common object counting approaches only predict the global objectcount and rely on additional instance-level supervision to also determineobject locations. We propose an image-level supervised approach that providesboth the global object count and the spatial distribution of object instancesby constructing an object category density map. Motivated by psychologicalstudies, we further reduce image-level supervision using a limited object countinformation (up to four). To the best of our knowledge, we are the first topropose image-level supervised density map estimation for common objectcounting and demonstrate its effectiveness in image-level supervised instancesegmentation. Comprehensive experiments are performed on the PASCAL VOC andCOCO datasets. Our approach outperforms existing methods, including those usinginstance-level supervision, on both datasets for common object counting.Moreover, our approach improves state-of-the-art image-level supervisedinstance segmentation with a relative gain of 17.8% in terms of average bestoverlap, on the PASCAL VOC 2012 dataset. Code link:https://github.com/GuoleiSun/CountSeg