 M. Yunchiao, a post-doctoral researcher from UIUC Image Formation and Processing Group, leading by Tom Swarn, this book provides a weekly supply solution for semantic sanitation. To relieve the demand for expensive pixel-lime annotations, weekly-supply information provides some promising solutions. Among these information, the simplest and the most efficient one can be collected for training is image lab labels. For this weekly-supply scheme, the key problem is how to build the relationship between image lab labels and pixels. Currently, several top-down attention projects are proposed to roughly localize objects with classification networks. We found that classifying networks are only responsive to small and sparse object regions. However, semantic sanitation requires dense object-laboring phosphoration. To extend object regions, we propose a very simple solution called Adversary Reaching. We first train a classifying network with original images, and CMS is employed to localize discriminative regions. Then we erase the main regions and retrain the network with the image, and then new object regions will be discovered. We continue this step for several times and merge the main regions from multiple steps as output. Such operation tries to challenge the classification network to discover some evidence of a specific category until no spot for evidence is left. Here is the problem that how many arrays the steps should be conducted. This figure shows the combination of classification training loss curves of different Adversary Reaching steps. We observe that loss value is very small when using original image for training. By performing a few steps to convert the loss value slightly increases. By continuing to perform one more step, the loss converges at a large value. We consider this state as our erasing. Our erasing may introduce many true negative regions into the main object regions. Therefore, we only use the main regions in the first three steps. We use silence detection techniques to produce background cues. The blue regions are no labelled pixels to employ those ignored regions as well as elevate noise. We further propose online PSL work to work with AE together. In particular, PSL introduced a classification branch. This branch provides classifying confidence to modulate the corresponding category-specific maps and form them into additional masks for separation. Those confidence maps with low classification confidences are needed for contributing to productive semantic semantics. We compare our methods with other state ODRs on PASCA, VLC 2012, and our master reached the best performance. In the future, we try to do some works like simultaneous weekly surprise object detection and semantic semantics or semi-surprised object detection and semantic semantics. Thanks.