FMD ( Full convolution MultiBox Detector ) is a simplified SSD ( https://github.com/weiliu89/caffe/tree/ssd ) with enabling different input sizes. It is just a learning project to master all details of a detector. Code is impllemented in Torch, is very simple and easy to understand.

• VGG16 convoluting image (224x224, or other sizes) to feature layer (14x14x512), just a typical full convolution network.

• Following feature layer, apply several kernels with different size and shape, such as 2x2, 2x3, 3x3, 2x4 ...

• Every kernel outputs a class score or a binding box according it's position (in feature layer) and kernel size. These kernels are just sliding windows with different postion and different size.

• Like SSD, not every predicted box do BP, just some selected ones, (positve and negtive is 1:4).

See model.lua, and boxsampleing.lua

In data folder, just see list.sh, filter.py, and makedb.lua. Combing all the info of VOC into a single t7 file.

See makefixed.lua, use fixed 10 layers from VGG16. Then see train.lua, it is very simple. The data batch is build in different threads ( see data_loader.lua and data_process.lua).

See doDetect.lua. The following is result of random pictures from web.