tag:blogger.com,1999:blog-2293611840857369720.post9120770347554436408..comments2024-03-27T11:47:51.024+03:00Comments on Computer Blindness: Object detection vs. Semantic segmentationAnonymoushttp://www.blogger.com/profile/01422787855469629259noreply@blogger.comBlogger5125tag:blogger.com,1999:blog-2293611840857369720.post-27435104549570780952018-11-02T13:30:40.011+03:002018-11-02T13:30:40.011+03:00Semantic segmentation may work like unsupervised c...Semantic segmentation may work like unsupervised classification of creation of clusters with some statistically significant boundaries and each cluster is supposed to represent few known labels. Each label could be object but in object detection one need to concentrate only on objects of interest to analyst Muralikrishna Iyyankihttps://www.blogger.com/profile/05414746466977328986noreply@blogger.comtag:blogger.com,1999:blog-2293611840857369720.post-63011112834321648962010-08-16T14:29:25.134+04:002010-08-16T14:29:25.134+04:00I see your point. While learning, the background c...I see your point. While learning, the background clutter adds noise to the learned distribution, if the bounding box is used. It could not be a problem on detection stage, if e.g. a bag-of-features approach is used, because it is enough to match only not-very-big rate of object features. Fair enough. But I don't think it is always true. If we want to use any shape cues, we definitely need to segment objects during detection. Also, the rule of thumb in ML is: the distribution of classes should be invariant of train/test stage.<br /><br />From the neuro-scientific point of view, the borders are detected on early stages, even before the brain -- in the eye retina. That's why this information is critical for detection in case of humans.<br /><br />You might also check out my follow-up post on this topic: http://computerblindness.blogspot.com/2010/08/image-parsing-unifying-segmentation.htmlAnonymoushttps://www.blogger.com/profile/01422787855469629259noreply@blogger.comtag:blogger.com,1999:blog-2293611840857369720.post-39620088671657265502010-08-03T16:50:22.674+04:002010-08-03T16:50:22.674+04:00I think, to robot or baby, 'Semantic segmentat...I think, to robot or baby, 'Semantic segmentation' will be processed for learning objects first and 'Object detection' will be processed for recognizing objects. What do you think?<br />I'm student from Korea.maun24https://www.blogger.com/profile/01948537604554173843noreply@blogger.comtag:blogger.com,1999:blog-2293611840857369720.post-91344770859390958442010-06-27T11:30:56.312+04:002010-06-27T11:30:56.312+04:00Fixed.
BTW, thank you for the feedback!Fixed.<br />BTW, thank you for the feedback!Anonymoushttps://www.blogger.com/profile/01422787855469629259noreply@blogger.comtag:blogger.com,1999:blog-2293611840857369720.post-847129975217977222010-06-23T00:19:22.587+04:002010-06-23T00:19:22.587+04:00"Semantic segmentation reduces easily to obje..."Semantic segmentation reduces easily to object detection" means "semantic segmentation can be solved if you have access to an oracle for the object detection task". You have meant the contrary here.<br /><br />http://en.wikipedia.org/wiki/Reduction_(complexity)hr0nixhttp://sexdrugsandappliedscience.comnoreply@blogger.com