VVC AUTO 85.5 - Smog Technician - Level 1 - Examples of Aftermarket Parts
Weakly Supervised Instance Segmentation using Class ...WILDCAT 53.5 PRM(Ours) 85.5 57.5 DCNNs...
1
Weakly Supervised Instance Segmentation using Class Peak Response Yanzhao Zhou 1 Yi Zhu 1 Qixiang Ye 1 Qiang Qiu 2 Jianbin Jiao 1 1 University of Chinese Academy of Sciences 2 Duke University Pattern Recognition and Intelligent System Development Laboratory http://ucassdl.cn 1 Motivation 3 Method 5 Application 6 Conclusion + Contact Cutting-edge instance segmentation methods require expensive rich annotated data for training deep models. Scan to get Paper & Code Yanzhao Zhou [email protected] http://yzhou.work 2 Our Goal A simple yet effective technique for weakly supervised instance segmentation using instance-aware cues from classification CNNs Enable classification networks for instance mask extraction horse Image-level labels are much cheaper and easier to define. Image Pixel-level Supervision Image-level Supervision Discover cues for each class and instance bird category representations Multi-instances are difficult to form a stable pattern Visual concepts within a single instance 4 Experiment Results Image PRM Prediction GT Typical failure The generation and utilization of Peak Response Maps Class Peak Response Peak Response Map We leverage class peak responses to extract fine-detailed instance-aware visual cues from classification networks PRM has great potential for bird FCN Classifier Image Peak Response Map C O N V person car Peak Backpropagation Peak Stimulation Loss C O N V 1 2 Training Inference C O N V Query Rank Prediction person person car NMS Segment Proposals Significant improvement in Pointwise Localization (mAP) Visualization shows clearer representations are learned via Peak Stimulation COCO VOC12 Method SPN 55.3 82.9 82.9 WILDCAT 53.5 85.5 PRM(Ours) 57.5 DCNNs trained using standard classification settings with negligible overhead High-quality instance-aware visual cues are obtained via Peak Backprop bird bird mAP SPN PRM(Ours) MELM 22.9 12.7 26.8 Method * 0.5 38.0 GTbbox potted plant potted plant
Transcript of Weakly Supervised Instance Segmentation using Class ...WILDCAT 53.5 PRM(Ours) 85.5 57.5 DCNNs...
-
Weakly Supervised Instance Segmentation using Class Peak ResponseYanzhao Zhou1 Yi Zhu1 Qixiang Ye1 Qiang Qiu2 Jianbin Jiao1
1 University of Chinese Academy of Sciences 2 Duke UniversityPattern Recognition and
Intelligent System Development Laboratoryhttp://ucassdl.cn
1 Motivation
3 Method
5Application
6Conclusion
+Contact
Cutting-edge instance segmentation methods require expensive rich annotated data for training deep models.
Scan to get Paper & CodeYanzhao [email protected]://yzhou.work
2 Our Goal
A simple yet effective technique for weakly supervised instance segmentation using instance-aware cues from classification CNNs
Enable classification networks for instance mask extraction
horseImage-level labels are much cheaper and easier to define.
Image Pixel-level Supervision
Image-level Supervision
Discover cues for each class and instance
bird category representations
Multi-instances are difficult to form a stable pattern
Visual concepts within a single instance
4 Experiment Results
Image PRM Prediction GT Typical failure
The generation and utilization of Peak Response Maps
Class Peak Response
Peak Response Map
We leverage class peak responses to extract fine-detailed instance-aware visual cues from classification networks
PRM has great potential
for bird
FCNClassifier
Image
Peak Response Map
CONV
person
car
Peak Backpropagation
Peak Stimulation
Loss
CONV
1
2
Trai
ning
Infe
renc
e
CONV
Query Rank
Prediction
personperson
car NMSSegment Proposals
Significant improvement inPointwise Localization (mAP)
Visualization shows clearer representations are learned via Peak Stimulation
COCOVOC12Method
SPN 55.382.982.9
WILDCAT 53.5
85.5PRM(Ours) 57.5
DCNNs trained using standard classification settings with negligible overhead
High-quality instance-aware visual cues are obtained via Peak Backprop
bird bird
mAPSPN PRM(Ours)MELM
22.9 12.7 26.8Method *
0.5 38.0GTbbox
potted plant
potted plant
Weakly Supervised Instance Segmentation using Class Peak Response�Yanzhao Zhou1 Yi Zhu1 Qixiang Ye1 Qiang Qiu2 Jianbin Jiao1�1 University of Chinese Academy of Sciences 2 Duke University
