Presentation is loading. Please wait.

Presentation is loading. Please wait.

Lecture 4a: Imagenet: Classification with Localization

Similar presentations


Presentation on theme: "Lecture 4a: Imagenet: Classification with Localization"— Presentation transcript:

1 Lecture 4a: Imagenet: Classification with Localization

2 Agenda ILSVRC competition Classification with Localization
Overfeat: integrated classification, localization, and detection R-CNN (Regions with CNN) SPP-net (Spatial Pyramid Pooling) Fast R-CNN

3 Imagenet Database Imagenet data base: 22 categories, ~14 mln labeled images ( ~700 images/class )

4 ILSVRC Classification over 1000 categories: Classification
Classification over 1000 categories: 1.2 million training images 50,000 validation images 150,000 testing images Classification Assign to each image label 5 guesses Classification & Localization 5 guesses: label + bounding box Detection: any number of objects in image (including zero) False positives are penalized

5 ILSVRC: Classification
top-5 labels

6 ILSVRC: Classification & Localization
top-5 labels + bounding box

7 ILSVRC 2014 vs PASCAL 2012 PASCAL 2012 ILSVRC 2013 ILSVRC 2014
# classes 20 200 Training # images 5,717 395,909 456,567 # objects 13,609 345,854 478,807 Validation 5,823 20,121 13,841 55,502 testing 10,991 40,152

8 ILSVRC: Classification
These are two dogs from two distinct classes

9 ILSVRC: Classification
Groundtruth: ????

10 ILSVRC: Classification
Groundtruth: coffee mug

11 ILSVRC: Classification
Groundtruth: coffee mug Top-5: table lamp lamp shade printer projector desktop computer

12 AlexNet (2012) AlexNet – winner 2012 with 85% top-5 accuracy
8 layers (5 conv. + 3 fully connected layers+droput + soft-max) 650K neurons , 60 Mln weights was trained on two GTX-580 with 3 GB memory. training took 6 days

13 AlexNet : Training SGD parameters batch = 128 examples,
momentum = 0.9, weight decay = Weight initialization: a 0-mean Gaussian with std dev= 0.01. Learning rate: The learning rate was initialized at 0.01 and adjusted manually throughout training: divide the learning rate by 10 when the validation error rate stopped improving The same for all layers, Dropout for fully connected layer 90 epochs through whole image dataset.

14 GoogleNet (2014) Winner of 2014 with 93.5% top-5 accuracy
22 layers (2 conv. +9 inception+ linear with dropout+ softmax) 5 mln parameters Trained on Google DistBelief cluster ~ 1 week on multi-GPU system (?)

15 Overfeat: Classification with localization

16 Overfeat: integrated classification, localization & detection
Overfeat (NYU) - a convolutional network to simultaneously classify, locate and detect objects. Key ideas: multiple scales apply a ConvNet at multiple locations in the image in a sliding window train the system to produce for each window a distribution over categories a prediction of the location and size of the bounding box with object relative to the viewing window accumulate results for each categories at each location and scale

17 Overfeat: topology summary
First 5 layers are similar to Alexnet: conv. layer with ReLU and max pooling, but with the following differences: no contrast normalization pooling regions are non-overlapping smaller stride to improve accuracy Input 3x [221x221], Output of last convoluitonal layer 1024 x [5x5] Feature Extraction: 3 x [231x231]  1024 x [5x5], down-sampling: 36:1

18 Overfeat: classification
Let’s takes image, and apply window [231x231]. Output of last conv. layer 1024 features x [5x5]. FC layers + log-loss will give scores for each class Feature extractor has multiple poolings and strides with total scale ratio 36:1, so if we slide the input window with step 36, the feature window will slide with step 1. 231x231 5x5 Input window [231x231] 1024 Features: x [5x5] 1000 class scores

19 Overfeat: classification
2 adjacent windows share many computations. Let’s get features for all windows locations simultaneously Feature Extraction: We compute first 5 layers for whole image ( correspond to 12:1 “subsampling” ) Classifier: The classifier has a fixed-size 5x5 input which is applied to the layer 5. We will shift the classifier’s input window by 1 pixel through pooling layers without subsampling. In the end we have [MxN] x C scores, where M, N are sliding windows index, and C – number of classes. Input Layer 5 Before pooling After pool 3x3 Classifier map 245x245 17x17 [3x3] x [5x5] [3x3] x C 281x 317 20x23 [6x9] x [5x5] [6x9] x C

20 Overfeat: data augmentation and scaling
to locate objects in different sizes we rescale image to 6 scales (The typical ratio from one scale to another is about ~1.4 ) horizontal flipping. Final post-processing: for each class we took local spatial max for different locations and scales take top-5 classes

21 Overfeat: Boosting Boosting: train 7 different models with different init weights, and select the best result

22 Overfeat: Training parameters
Data augmentation: Each image is down-sampled so that the smallest dimension is 256 pixels. Then extract 5 random crops (and their horizontal flips) of size 221x221 pixels Training: weight initialization: randomly with (µ, σ) = (0, 1 × ) SGD with momentum =0.6 weight decay =1×10-5 learning rate = 0.05 , is decreased by ½ after 30, 50, 60, 70, 80 epochs dropout

23 Overfeat: Localization
Starting from classification-trained network Fix the feature extraction layers (1-5) and replace the classifier layers by a regression network Regression net takes as input the features from layer 5 2 fully-connected hidden layers of size 4096 and 1024 channels output layer 4 units for each class: with the coordinates for the bounding box edges. Train regression net the same set of scales as in multi-scale classification ℓ2 loss between the predicted and true bounding box for each example compare the prediction of the regressor at each spatial location with the ground-truth bounding box, shifted into the frame of reference

24 Overfeat: Regression Net Topology
Input 3x231x231 conv: 11×11 stride 4×4; ReLU; maxpool: 2×2 stride 2×2; output: 96x24x24 conv: 5×5 stride 1×1; ReLU; maxpool: 2×2 stride 2×2; output: 256x12x12 conv: 3×3 stride 1×1 0-padded; ReLU; output: 512x12x12 conv: 3×3 stride 1×1 0-padded; ReLU; output: 1024x12x12 conv: 3×3 stride 1×1 0-padded; ReLU; maxpool: 2×2 stride 2×2; output: 1024x6x6 full; ReLU; output : 4096x1x1 full; output: 1024x1x1 Regressor: output: 200 x 4

25 Overfeat: localization pipeline
The raw classifier/detector outputs a class confidence for each window location (multi- scale)

26 Overfeat: localization pipeline
The regression predicts the location of the object with respect to each window (multi-scale)

27 Overfeat: localization pipeline
3. Box merging for top5 classes

28 Overfeat: Localization pipeline
Choose top-5 classes Cs by taking the maximum detection class outputs across spatial locations for each scale s ∈ 1… 6 Initial set of bounding boxes: B ←Us Bs, where Bs the set of bounding boxes predicted by the regressor net for each class in Cs, across all spatial locations at scale s. Repeat merging of boxes from B until done: (b1, b2) = argmin b1!= b2∈B match_score (b1, b2) , where match_score = the sum of the distance between centers of the two bounding boxes and the intersection area of the boxes. If (match_score(b1, b2) < t), then set B ← B\ {b1, b2} ∪ box_merge(b1, b2): box_merge computes the average of the bounding boxes’ coordinates.

29 R-CNN: Regions with CNN

30 R-CNN: Regions with CNN features
Regions with CNN detection approach: generates ~2000 category-independent regions for the input image, extracts a fixed-length feature vector from each region using a CNN, classifies each region with category-specific linear SVM R-CNN outperforms OverFeat, with a mAP = 31.4% vs 24.3%. R. Girshick et al , “Rich feature hierarchies…”

31 R-CNN pipeline Region detection  2000 regions , see Feature extraction with Imagenet: Region croped and scaled to [227 x 227] 5 conv.layers + 2FC  4096 features SVM for 200 classes Greedy non-maximum region suppression for each class: rejects a region if it has an big overlap with a region which has > score

32 R-CNN Training The key idea is to train feature extraction CNN on a large auxiliary dataset (ILSVRC-classification 1000 classes), followed by domain specific fine-tuning on a small dataset (200 classes): Pre-training: Train Imagenet for classification Replace last layer with FC layer to N+1 outputs: N classes + 1 “background” if ground truth class and gt box has IoU with region > ½  positive, otherwise  background Regular SGD based training batch from 128 images = 32 positive + 96 background Random weights initialization

33 R-CNN: ILSVRC 2013 detection performance

34 R-CNN drawbacks Training is expensive in time:
Training is a multi-stage pipeline. Fine-tune a ConvNet for detection using cross-entropy loss. Train linear SVMs on ConvNet features computed on warped object proposals Learn bounding-box regressors Training is expensive in storage: For SVM and regressor training, features are extracted from each warped object proposal in each image and written to disk – hundreds of GB Test-time detection is slow.

35 R-CNN speed and R-CNN detection is very slow: 12.5 sec/image
time per 1 frame analysis

36 CNN with Spatial Pyramid pooling

37 SPP-net = Spatial Pyramid Pooling + CNN
Classical conv. net takes a fixed-size (e.g ) input image: Need cropping or warping to transform original image to square shape This constraint is related to Fully-Connected layer ONLY Idea: let’s use Spatial Pooling Pyramid to transform any-shape image to ‘fixed-length” feature vector. Kaiming He et al, “Spatial Pyramid Pooling in Deep Convolutional Networks …“

38 SPP-Net Soft Max Inner Product BACKWARD FORWARD ReLUP Inner Product
SPP(5x5+7x7+13x13) Pooling [2x2, stride 2] Convolutional layer [5x5] Pooling [2x2, stride 2] Convolutional layer [5x5] Data Layer Data Layer [28x28]

39 Spatial Pyramid Pooling
Example. SPP layer : Input feature maps has size [13 x13] 3 pooling layers, which have fixed outputs size: [4x4], [2x2], and [1x1] Size of pooling windows and stride for each pooling layer depends on input feature map size.

40 SPP-net training Network - based on Overfeat Training:
Data augmentation: horizontal flipping and Color altering Dropout with 2 last FC layers Init learning rate =0.01; divide by 10 when error plateau

41 SPP-net: Detection pipeline
find 2000 region-candidates ( like R-CNN ) extract the feature maps from the entire image SPP maps each window of the features, corresponding to region-candidate, to a fixed-length representation SVM using 2 FC layers

42 SPP-net vs R-CNN Detection is ~100x faster than R-CNN, but training is very slow: Multi-stage training (fine-tuning of last layers, SVM and regressors) Still needs a lot of disk space to save features vectors

43 Fast R-CNN

44 Fast-RCNN Fast R-CNN is similar to SPP-net, but it trains both classification and regression networks together using a multi-task loss in a single training stage: Unlike SPP-net, all network layers can be updated during fine-tuning The multi-task loss simplifies learning and improves detection accuracy Currently the best and the fastest detection algorithm 3x faster training, Detection only 0.3s , Best performance R. Girshick, Fast R-CNN: code:

45 Fast-RCNN: architecture
Start from CNN trained for ImageNet classification Add Region-of-Interest (RoI) pooling layer after last convolutional layer Replace FC layers by two ‘sibling” nets: Sof-max: estimates probability over (K+1 background) classes Regression layer which predicts box (x,y,w,h) for each K classes

46 Fast-RCNN: architecture
Each RoI is pooled into fixed size vector 2 FC layers  RoI feature vector Multi-task loss: softmax classifier + box regressor/ class

47 Fast-RCNN: RoI pooling layer
Input: feature maps from the last convoluional layer [C x H x W] List of regions of interest . Each RoI is a tuple (n; x; y; h;w) where n - image index (or scale), (r; c) region top-left location and (h;w) - height and width Output: for each region Max-pooled feature maps [C x H’ x W’] (H’ ≤ H, d W’ ≤ W). The pooling area (“bin” ) size ~ [h/H’, w/W’] ( SPP with one level) RoI pooling layer back-propagation:

48 Fast-RCNN: Multi-task Loss
For each RoI 2 outputs: Sof-max: estimates probability over (K+1 background) classes Regression layer: predicts box b=(x,y,w,h) for each class Multi-task loss: L(p, k*, b, b*) = Lcls (p, k*) + λ*δ(k*) Lloc (b,b*) k* is true label: δ(k* ) =1 and 0 for all other class Lcls is regular log-loss: Lcls (p, k*) = -log (pk*) Lloc is smooth L1 loss: Lloc (b,b*) = [ x-x’ ] + [ y-y’ ] + [ h-h’ ] +[ w-w’ ], where [t] is smooth L1 norm: [t] = min ( 0.5*|t| , 0.5*t2 )

49 Fast-RCNN: One pass training
Training is combines together fine-tuning of convolutional layers, training of of classifier and box-regressors: Each batch consists of 2 images x 64 RoI per image: 16 “true” RoI ( IoU > ½ with ground truth box) 48 ‘background” RoI ( IoU < ½) , labeled as k=0 Data augmentation: each image is horizontally flipped with prob= ½ Option: multi-scale image pyramid SGD: Momentum =0.9, weight decay=0.0005 weight initialization for FC layers – gaussian (σ = 0.01, 0.001) initial global lr =0.001, after 30,000 iterations lr =0.0001

50 Fast R-CNN: detection time optimization
For classification FC layers usually take small amount of time comparing to convolutional layers. For detection FC will become heavy compute since we have to do it for each RoI. We can compress a single fully-connected layer W using SVD-like decomposition of W = U*∑*VT into two fully-connected layers without a non-linearity between them: First FC corresponds to (U*∑) and second to VT

51 BACKUP

52 ILSVRC: Localization

53 Overfeat: classification

54 Overfeat: ”fast” net topology
Input 3x231x231 convo: 11×11 stride 4×4; ReLU; maxpool: 2×2 stride 2×2; output: 96x24x24 convo: 5×5 stride 1×1; ReLU; maxpool: 2×2 stride 2×2; output: 256x12x12 convo: 3×3 stride 1×1 0-padded; ReLU; output: 512x12x12 convo: 3×3 stride 1×1 0-padded; ReLU; output: 1024x12x12 convo: 3×3 stride 1×1 0-padded; ReLU; maxpool: 2×2 stride 2×2; output: 1024x6x6 convo: 6×6 stride 1×1; ReLU; output: 3072x1x1 full; ReLU; output : 4096x1x1 full; output: 1000x1x1 softmax; output: 1000x1x1

55 Single-class Regression vs Per- Class Regression
Using a different top layer for each class in the regressor network for each class (Per-Class Regressor (PCR) surprisingly did not outperform using only a single network shared among all classes (44.1% vs. 31.3%).

56 Overfeat: Detection The detection task differ from localization in that there can be any number of object in each image (including zero), and that false positives are penalized by the mean average precision (mAP) measure The main difference with the localization task, is the necessity to predict a background class when no object is present. Traditionally, negative examples are initially taken at random for training. Then the most offending negative errors are added to the training set in bootstrapping passes.

57 R-CNN: PASCAL VOC performance
2012 SIFT, HOG,…

58 R-CNN: PASCAL VOC performance
2014: Regions with CNN

59 R-CNN CODE Requires Matlab!

60 CNN regression Szegedy et all ( Google) 2010, “Deep Neural Networks for Object Detection” start with Alexnet, replace last soft-max layer with regression layer which generates an binary mask “d x d” : 1 if pixel is inside box, 0- otherwise; train net by minimizing L2 error vs ground truth mask m:

61 CNN regression Multi-scale

62 CNN regression Issues: Issue1:
Overlapping masks for multiple touching objects Localization accuracy Recognition of small objects Issue1: To deal with multiple touching objects, we generate not one but several masks, each representing either the full object or part of it. we use one network to predict the object box mask and four additional networks to predict four halves of the box: bottom, top, left and right halves


Download ppt "Lecture 4a: Imagenet: Classification with Localization"

Similar presentations


Ads by Google