logoRNN

CMPT 414- CV Modeling Term Project

Our project focuses on implementing region based CNN for logo detection

Object Proposal Generator

Dependencies

This project requires OpenCV (and Caffe for Fast RCNN)

Build Object Proposal Generator

I recommend building in a separate directory

mkdir build && cd build

Generate makefiles and build from inside the build directory,

cmake <path/to/logoRNN> && cmake --build

Note that the project requires third party library (egbis) found on github and building will attempt pull and merge it to local. Building offline will fail.

Components

For demonstration purposes, building creates 3 executables:

opgen
optest
viewer

The executables are found in directory ./bin/bin.

How to Run

Opgen

./bin/bin/opgen <path/to/image> colorweight textureweight sizeweight fillweight topk

weight and topk arguments are optional and will default to 1 for weights and 3 for topk

Run the generator to get the top right and bottom right corner of the topk object proposals

Each line will be 4 numbers separated by commas in the format top left x, top left y, bottom right x, bottom right y.

Running the generator on ./tests/imgs/test1.jpg will yield the following:

./tests/imgs/test1.jpg 136 0 293 76
./tests/imgs/test1.jpg 217 0 499 231
./tests/imgs/test1.jpg 217 0 499 231

Optest

./bin/bin/optest <path/to/image> colorweight textureweight sizeweight fillweight

Similar to Opgen, Optest's weight arguments are optional and default to 1.

Run the test to obtain the following window:

Viewer

./bin/bin/viewer <path/to/image> colorweight textureweight sizeweight fillweight

Run the viewer to obtain a set of windows demonstrating each iteration(level) of grouping:

Warning: Viewer may generate a lot of windows for highly segmented images, I recommend using test images instead

Experiment

Our scoring test will generate a text file describing how object proposal accuracy for each combination of weight parameters with each parameter being either 0 or 1.

Build Opgen
Unzip data
Run scoringprocess.sh

FUTURE PLANS -- INCOMPLETE --

RCNN setup

Build opgen
Unzip data
run train_proposals.sh and test_proposals.sh to generate a list of bounding boxes
Git pull submodule caffe-fast-rcnn

Training

With GPU

./pythontools/train_net.py --gpu 0 --solver models/solver.prototxt --weights data/models/logos.model

Without GPU

./pythontools/train_net.py --cpu-only --solver models/solver.prototxt --weights data/models/logos.model

Testing

Yet to be complete

mingkaic/logoRNN