NoListen / mxnet-reinspect

mxnet-reinspect

MXNet version of Reinspect

Features

• No training included
• 35 fps for detection (Nvidia TITANX)
• enlarge the bbox to the whole body by scales varying from scenes to scenes
• ROI Pooling for bounding boxes to extract features

Additional Requirements

• Transfer caffemodel to MXNet using caffe_converter in mxnet/tools.

Illustration

• mxnet_track.py is the main file
• config.json describes the hyperparameters
• reinspect.json is the network architecture file
• utils is similar to that in Reinspect with redundant files moved.
• model-transfer specifies the needs to deal with the model learned in Reinspect

Process

• mxnet model load googlenet params
• mxnet model load lstm params (lstm.h5)
• output the proposals
• extract the features of bbox proposals

Note

• tracking is not included which you can refer to MOT_XCODE.
• bridging python and C++ is not included which you can refer to numpy-opencv-converter.

TODO

• add learning process to mxnet-reinspect.