Image Segmentation and CNN Implementation to Detect Humans We tried three approaches. First, we tried to use a native C++ model and adapt it to our dataset. Its broken. Second, we made a model in python, exported it to C++ tensorflow and it works fast. Really fast. Third, we tried to compare that with our original model in python, which is the python script installed. For python, we used the YOLOv1 algorithm, which worked better than our C++ box counterpart.
- Native CNN: a. CMAKE b. Armadillo c. Boost
- C++ Model a. Bazel b. Tensorflow (use the tensorClone.sh shell script) c. openCV note- you have to clone tensorflow into this repo, then mv WORKSPACE and the cpp model to tensorflow/cc/examples otherwise, it wont run.
- Python a. Tensorflow b. Keras c. Scipy d. PIL e. Yad2k
The dataset that we used for most of the training until Austin finished his data compilation is here. Download it and move the files into this directory. http://pascal.inrialpes.fr/data/human/
Austin, please put the link to our total dataset here
For each model, run the following:
./[insert model exec file name] --image
- fix native CNN implementation
- make it so you dont have to clone TF
- Fix python bugs on runtime