SOLO ：简单高效的视觉实例分割和识别框架.pdf

1、SOLO:SegmentingObjects byLocationsTaoKong孔涛http:/www.taokong.orqResearcher，ByteDance AI LabllByteDance字节跳动Joint work with Xinlong Wang. Rufeng Zhang，Yuning Jiang， Lei Li and Chunhua Shen#page#Outline Background and Introduction Previous Instance Segmentation Solutions SOLO (ECCV-2020） & SOLOv2 (Neur

2、iPS-2020）SummaryllByteDance字节跳动#page#Introduction: Visual Perception Problems Object Detectionxoq Bulpunoq e Buisn uoea azjeool pue spoolqo AISSE ol jeo5Output:A semantic label and a bounding box for each instance.Output space: N*(4+C） Semantic Segmentation：Goal: To classify each pixel into semantic

3、 categoriesOutput: A semantic label for each pixelOutput space:H*W*C后动 Instance Segmentation:Goal: To classify objects and localize each using a mask.Output:A semantic label and a mask for each instance.门Output space:？Figure crediit: He et al 2017#page#Introduction: Instance Segmentation is HARD to

4、formulateInstance-level prediction & pixelprediction Instance may occur at each position! Occlusions! Poses, scales and deformations!Need fine boundary！lByteDance字节跳动#page#Introduction: Instance Segmentation ApplicationsIntelligent DrivingllByteDance字节跳动#page#Introduction: Instance Segmentation Appl

5、icationsIntelligent RobotInput Point CloudsInstance Seamentation ResultsllByteDance字节跳动#page#Introduction: Instance Segmentation ApplicationsIntelligent EditingllByteDance字节跳动#page#Introduction: Instance Segmentation ApplicationsIntelligent Life&WorklByteDancee字节跳动#page#Previous SolutionsDeep Learni

6、nge.g.， Normalized Cuts(2000）e.g，Mask R-CNN(2017）2015年Important but ignored due to time limitationWill be discussed laterllByteDance字节跳动#page#SOLOv2 results (ResNet-101，V100 GPU）#page#Previous Solutions: two methodsRPNFCNGroupingDetect-then-segment，Label-then-cluster，eg Discriminative loss,SGN,SSAP，