1y ago

30 Views

1 Downloads

1.87 MB

56 Pages

Transcription

Fast R-CNNRoss GirshickFacebook AI Research (FAIR)Work done at Microsoft ResearchPresented by:Nick Joodi Doug Sherman

Fast Region-based ConvNets (R-CNNs)FastSorry about the black BG, Girshick’s slides were all black.2

The Pascal Visual Object Classes Challenge Overview Classification, Detection, SegmentationFor each image: Does it contain the class? classification Where is it? detection via bounding boxEvaluation Mean Average Precision (mAP) Participants submitted results in the form ofconfidence Produce Precision Recall curves Average precision for each class Take mean to get mAP3

Object detection renaissance (2013-Present)4Adapted from Fast R-CNN [R. Girshick (2015)]

Object detection renaissance (2013-Present)5Adapted from Fast R-CNN [R. Girshick (2015)]

Object detection renaissance (2013-Present)6Adapted from Fast R-CNN [R. Girshick (2015)]

Agenda1.Pre-existing Modelsa.b.2.Ways to improvea.b.3.“Slow” R-CNNSPP-netSGD Mini-BatchNew Loss FunctionFast R-CNNa.b.ArchitectureResults & Future Work7

Region-based convnets (R-CNNs) R-CNN (aka “slow R-CNN”) [Girshick et al. CVPR14]SPP-net [He et al. ECCV14]8

9Adapted from Fast R-CNN [R. Girshick (2015)]

10Adapted from Fast R-CNN [R. Girshick (2015)]

11Adapted from Fast R-CNN [R. Girshick (2015)]

12Adapted from Fast R-CNN [R. Girshick (2015)]

13Adapted from Fast R-CNN [R. Girshick (2015)]

14Adapted from Fast R-CNN [R. Girshick (2015)]

What’s wrong with slow R-CNN? Ad hoc training objectives Fine-tune network with softmax classifier (log loss)Train post-hoc linear SVMs (hinge loss)Train post-hoc bounding-box regressors (L2 loss)15

What’s wrong with slow R-CNN? Ad hoc training objectives Fine-tune network with softmax classifier (log loss)Train post-hoc linear SVMs (hinge loss)Train post-hoc bounding-box regressors (L2 loss)Training is slow (84h), takes a lot of disk space16

What’s wrong with slow R-CNN? Ad hoc training objectives Fine-tune network with softmax classifier (log loss)Train post-hoc linear SVMs (hinge loss)Train post-hoc bounding-box regressors (L2 loss)Training is slow (84h), takes a lot of disk spaceInference (detection) is slow 47s / image with VGG16 [Simonyan & Zisserman. ICLR15]Fixed by SPP-net [He et al. ECCV14]17

Agenda1.Pre-existing Modelsa.b.2.Ways to improvea.b.3.“Slow” R-CNNSPP-netSGD Mini-BatchNew Loss FunctionFast R-CNNa.b.ArchitectureResults & Future Work18

19Adapted from Fast R-CNN [R. Girshick (2015)]

20Adapted from Fast R-CNN [R. Girshick (2015)]

21Adapted from Fast R-CNN [R. Girshick (2015)]

22Adapted from Fast R-CNN [R. Girshick (2015)]

23Adapted from Fast R-CNN [R. Girshick (2015)]

24Adapted from Fast R-CNN [R. Girshick (2015)]

Pyramid Pooling Layer(w/4 x h/4)(2 x 1)To FC(w/2 x h/2)(4 x 2)812341(w/1 x h/1)(8 x 2)2344RegionStride/WindowSizeOutput ofPoolingConcatenated25

26Adapted from Fast R-CNN [R. Girshick (2015)]

What’s wrong with SPP-net? Inherits the rest of R-CNN’s problems Ad hoc training objectiveTraining is slow (25h), takes a lot of disk spaceIntroduces a new problem: cannot update parameters below SPP layerduring training27

28Adapted from Fast R-CNN [R. Girshick (2015)]

Agenda1.Pre-existing Modelsa.b.2.Ways to improvea.b.3.“Slow” R-CNNSPP-netSGD Mini-BatchNew Loss FunctionFast R-CNNa.b.ArchitectureResults & Future Work29

SGD Mini-Batch Method for RoIs30Adapted from Fast R-CNN [R. Girshick (2015)]

SGD Mini-Batch Method for RoIs31Adapted from Fast R-CNN [R. Girshick (2015)]

SGD Mini-Batch Method for RoIsInput size for SPP-net32Adapted from Fast R-CNN [R. Girshick (2015)]

SGD Mini-Batch Method for RoIs33Adapted from Fast R-CNN [R. Girshick (2015)]

SGD Mini-Batch Method for RoIs34Adapted from Fast R-CNN [R. Girshick (2015)]

SGD Mini-Batch Method for RoIs35Adapted from Fast R-CNN [R. Girshick (2015)]

SGD Mini-Batch Method for RoIs36Adapted from Fast R-CNN [R. Girshick (2015)]

SGD Mini-Batch Method for RoIs37Adapted from Fast R-CNN [R. Girshick (2015)]

Agenda1.Pre-existing Modelsa.b.2.Ways to improvea.b.3.“Slow” R-CNNSPP-netSGD Mini-BatchNew Loss FunctionFast R-CNNa.b.c.ArchitectureResultsFuture Work38

Revised loss functionFor the classificationFor the bounding boxp: Predicted RoI Classificationu: True RoI Classificationtu (tx,ty,tw,th): Predicted Bounding Boxv (vx,vy,vw,vh): True Bounding Boxƛ : Controls the balance between the two losses39

Revised loss function40

Revised loss functionSmooth: Continuously Differentiable41

Agenda1.Pre-existing Modelsa.b.2.Ways to improvea.b.3.“Slow” R-CNNSPP-netSGD Mini-BatchNew Loss FunctionFast R-CNNa.b.ArchitectureResults & Future Work42

Fast R-CNN Fast test-time, like SPP-netOne network, trained in one stageHigher mean average precision than slow R-CNN and SPP-net43

44Adapted from Fast R-CNN [R. Girshick (2015)]

45Adapted from Fast R-CNN [R. Girshick (2015)]

46Adapted from Fast R-CNN [R. Girshick (2015)]

47Adapted from Fast R-CNN [R. Girshick (2015)]

48Adapted from Fast R-CNN [R. Girshick (2015)]

49Adapted from Fast R-CNN [R. Girshick (2015)]

50Adapted from Fast R-CNN [R. Girshick (2015)]

Agenda1.Pre-existing Modelsa.b.2.Ways to improvea.b.3.“Slow” R-CNNSPP-netSGD Mini-BatchNew Loss FunctionFast R-CNNa.b.ArchitectureResults & Future Work51

52Adapted from Fast R-CNN [R. Girshick (2015)]

53Adapted from Fast R-CNN [R. Girshick (2015)]

54Adapted from Fast R-CNN [R. Girshick (2015)]

What’s still wrong? Out-of-network region proposals Selective search: 2s / img; EdgeBoxes: 0.2s / imgFortunately, this has already been solvedS. Ren, K. He, R. Girshick & J. Sun. “Faster R'CNN: Towards Real'TimeObject Detection with Region Proposal Networks.” NIPS (2015).55

Fast R-CNN take-aways End-to-end training of deep ConvNets for object detectionFast training timesOpen source for easy experimentationA large number of ImageNet detection and COCO detection methods arebuilt on Fast R-CNN56

Fast R-CNN a. Architecture b. Results & Future Work Agenda 42. Fast R-CNN Fast test-time, like SPP-net One network, trained in one stage Higher mean average precision than slow R-CNN and SPP-net 43. Adapted from Fast R-CNN [R. Girshick (2015)] 44.

Related Documents: