Unsupervised Object Localization: Observing the Background to Discover Objects

Figure 1 . Examples of object localization results obtained with our method FOUND on images from diverse datasets. We propose a simple framework in which we train a single conv1 × 1 layer, and achieve state-of-the-art results in unsupervised object discovery and saliency detection. We train for only 2 epochs over the 10k dataset DUTS-TR [54] and inference runs at 80 FPS. Note that the results presented here are without post-processing refinement.