OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks

Key ideas

Multiscale and sliding window can be implemented within a ConvNet
Learning to predict object boundaries
Bounding boxes are then accumulated instead of supressed to increase confidence
Winner or ILSVRC2013 and obtained competitive results for detection and classification
Feature extractor from OverFeat

Main advantage of CNN is that you train from raw pixels to ultimate categories: thereby alleviating the requirement to design feature extractors
Main disadvantage of CNN is that you need a lot of labeled training samples
Point of the paper: show that training CNN to classify, locate and detect objects can boost the accuracy of all 3
While ImageNet usually has the main object roughly centered, objects of interest may vary in size and position
- Address this with CNN + sliding window over multiple scales - however many windows contain an identifiable portion of object but not the whole object
- Instead of producing distribution of categories for each ewindow, produce a prediction of location and size of bounding boxes
Other autors proposed to classify the central pixel, then categorize using the window as context for the decision
Order of difficulty: classification < localization < detection

Similar to best ILSVRC12 by Krizhevsky et al
Fixed input size during training, smallest dimension is 256 pixels
Extract 5 random crops and horizontal flips of size 221x221, present them in mini-batches of size 128
Feature extractor: OverFeat to provide powerful features for computer vision research. Fast and accurate model provided.
Multi-view voting to boost performance: Fixed set of 10 views (4 corners + center with horizontal flip) is averaged
- While sliding window might be prohibitive for certain types of models, for CNN it's efficient
- Output will be a spatial map of C dimensional vectors at each scale
Why are sliding windows eficient for CNNs? Because they naturally share computations common to overlapping regionso

Replace classification layers by regression network
Simultaneously run classifier and regression networks across all locations and scales, since the feature extractor layers are the same, only the final FC layers have to be recomputed
Input to regression network: pooled feature maps from layer 5, 2 FC layers of 4096 and 1024, final output layer of 4 units for BBox edges coordinates
Combining predictions by greedy merge strategies
max_score is computed with the sum of the distance between centers of the two bounding boxes and intersection area of the boxes