[Paper note] PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection

paper

Author: Kye-Hyeon Kim, Sanghoon Hong, Byungseok Roh, Yeongjae Cheon, and Minje Park

Intel Imaging and Camera Technology

Highlight

Speed up (real-time) detection process with a more efficient feature extraction CNN, without lossing too much accuracy.

This network is smaller and more efficient than ResNet and can be a substitution of the latter.

Main structure

C.ReLU: Concatenated ReLU in early stage of CNN to reduce the number of computation.

In early stages, output nodes tend to be paired, i.e. one node’s activation is the opposite of another’s.

C.ReLU reduce the output channels by half and concatenate with negation.

Inception

Not yet been widely applied.

Cost-efficient building block for multi-scale representation.

1x1 Conv can preserve the receptive field of the previous layer.

Multi-scale representation like HyperNet

Concatenate the output of the last layer and two intermediate layers, whose size are 2x and 4x of the last layer.

Set 2x layer as reference scale, down-scaling (pooing) 4x layer, up-scaling (interpolation) the last layer.

Experiment

Training details

Add Batch normalization layers before all ReLU.

Plateau detection based learning rate policy.

Measure the moving average of loss

Decide as on-plateau if its improvement is below a threshold.

Decrease the learning rate by a constant factor when on-plateau.

Number of proposals = 200

VOC2007 & VOC2012 performance