Experiments with learning depth from a single image for course project.
Details and history in presentation here: https://docs.google.com/presentation/d/1GZkUPq5xFVHfjKiVrPzIXjKb_lBb_r4KPjPVz_VMwK8/edit?usp=sharing
Example Results:
I used data from the Make3d Dataset found here: http://make3d.cs.cornell.edu/data.html
I organized the data as follows: The test and training data were downloaded to a directory called "data" and then into subdirectories:
../data/test/depthmaps
http://www.cs.cornell.edu/~asaxena/learningdepth/Data/Dataset2_Depths.tar.gz
../data/test/images
http://cs.stanford.edu/people/asaxena/learningdepth/Data/Dataset2_Images.tar.gz
../data/train/depthmaps
http://cs.stanford.edu/people/asaxena/learningdepth/Data/Dataset3_Depths.tar.gz
../data/train/images
http://www.cs.cornell.edu/~asaxena/learningdepth/Data/Dataset3_Images.tar.gz
After downloading and extracting the data to the appropriate paths, I rotated the images (they are originally flipped 90 degrees from the depthmaps and resized them to make the model train more quickly with the following commands:
sh resize_and_rotate.sh ../data/train/images/ ../data/train/small_images/
sh resize_and_rotate.sh ../data/test/images/ ../data/test/small_images/
conda env create -f mono_environment.yml
To run with debug statements on cpu:
THEANO_FLAGS="device=cpu,optimizer=None,compute_test_value=raise,floatX=float32" python predict_depth.py
To run on cpu:
THEANO_FLAGS="device=cpu,floatX=float32" python predict_depth.py
To run on gpu with no debug:
THEANO_FLAGS="device=gpu,floatX=float32" python predict_depth.py
Error in bottom few rows of pixels