5 Why Weakly Supervised Localization (WSL)? Knowing where to look, recognizing objects will be easier !However, in the classification-only task, no annotations of object location are available.Weakly Supervised Localization
7 13.9: Weakly supervised object detector learning with model drift detection, ICCV 2011 15.0: Object-centric spatial pooling for image classification, ECCV 201222.4: Multi-fold mil training for weakly supervised object localization, CVPR 201422.7: On learning to localize objects with minimal supervision, ICML 201426.2: Discovering Visual Objects in Large-scale Image Datasets with Weak Supervision, submitted to TPAMI26.4: Weakly supervised object detection with posterior regularization, BMVC 201431.6: Weakly supervised object localization with latent category learning, ECCV 2014Sep 11, Poster Session 4A, #34
8 Our WorkVOC 2007ResultsOurs31.6DPM 5.033.7VOC 2007ResultsOurs26.2DPM 5.033.7Weakly Supervised Object Localization with Latent Category LearningDiscovering Visual Objects in Large-scale Image Datasets with Weak SupervisionECCV 2014Submitted to TPAMIFor the consideration of high efficiency in large-scale tasks, we use the second one.
18 3rd: Detection Rescoring Rescoring with softmaxtrainsoftmaxmax……128 boxes…………1000 dim1000 dim1000 classesSoftmax: consider all the categories simultaneously at each minibatch of the optimization – Suppress the response of other appearance similar object categories
19 4th: Classification Rescoring Linear Combination………1000 dim1000 dim1000 dimOne funny thing: We have tried some other strategies of score combination, but it seems not working !
24 2nd: MILinear on ILSVRC 2013 detection mAP: 9.63%! vs 8.99% (DPM5.0)
25 2nd: MILinear for Classification MethodsTop 5 ErrorMilinear17.1
26 3rd: WSL Rescoring (Softmax) MethodTop 5 ErrorBaseline with one CNN :13.7Average with four CNN :12.5MILinear17.1MILinear + Rescore13.5The Softmax based rescoring successfully suppresses the predictions of other appearance similar object categories !
27 4th: Cls and WSL Combinataion MethodTop 5 ErrorBaseline with one CNN model:13.7Average with four CNN models:12.5MILinear17.1MILinear + Rescore13.5Cls (12.5) + MILinear (13.5)11.5WSL and Cls can be complementary to each other!
28 Russakovsky et al. ImageNet Large Scale Visual Object Challenge.
29 Conclusion WSL always helps classification WSL has large potential: WSL data is cheap