ys0129 / oirds Goto Github PK

(1) For python scrips to run: Changed import order/statement prior to import of cafee: import matplotlib matplotlib.use('Agg')

(2) File layout: 1. Load the metadata from .xls files. - gen_file_lists.py 2. Correct for ground sample distance variation and modify the centroid and polygon coordinates. - tools/prep.py; tools/img_manip.py 3. Remove vehicles with shadows. - tools/prep.py 4. Rotate the images in 15 degree increments and modify the centroid and polygon coordinates. - tools/prep.py; img_manip.py 5. Select patches of a given size around vehicles in the images. - tool/prep.py 6. Crop a 'no vehicle' patches from images with only one vehicle. - tools/prep.py 7. Create the corresponding training and validation set .txt files with a user defined train-test ratio. - tools/prep.py 8. Label OIRDS images by pixel (and plot the training images), as required for Fully Convolutional Networks for Semantic Segmentation. - tools/label_image.py 9. Convert .tif files to .png using ImageMagick. - tools/convert_to_png.sh 10. Print output from the models via google logs. - tools/glogs.py 11. Generate the Lightning Memory-mapped Databases (train and test) and training image mean. - tools/warm_up.py 12. Achieve sweet accuracy with CNN on cropped images. - 'caffe train -solver models/small_solver.prototxt -gpu 0' 13. Fine-tune a CNN on OIRDS. - tools/finetuning/oirds-fine-tuning.py

*0. Install instructions and sample prototxts for Fully Convolutional Networks for Semantic Segmentation.
    - tools/fcn/
*1. Tools for Region-based (sliding window) CNNs
    - tools/rcnn/
*2. Describe the environment of a successful installation.
    - describe.sh; description.txt

F-CNN Work: 1. Convert images to png files. Place in png directory with top level Data Set. Last argument is the fixed label index for cars. 2. python create_train_dbs.py create_db.json -- creates label and data LMDB files (raw_train, raw_test, groundtruth_train, groundtruth_test) -- stores images in png_raw and png_gt (cropped if resized) -- images are assumed to be in png directory under 'dataDirectory' 3. /opt/fcn/caffe/build/tools/compute_image_mean raw_train train_mean.binaryproto -- skip step if not resizing 4. /opt/fcn/caffe/build/tools/compute_image_mean raw_test test_mean.binaryproto -- skip step if not resizing 5. /opt/fcn/caffe/build/tools/caffe train -solver fcn8_solver.prototxt -weights fcn-8s-pascalcontext.caffemodel -gpu 0 -- the model comes from https://gist.github.com/longjon/1bf3aa1e0b8e788d7e1d#file-readme-md
OR python solver_with_netsurgery.py solver.json 6. python test_label.py test_label.json 0.1 -- tests a random sample of images (1% of the data) from 'dataDirectory' data set. The images are assumed to be in a directory called 'png'. T -- This python routine dumps a stats.txt file containing tuples for each image: (x dim, y dim, false positive, false negative, true positive, true negative, wrong label, precision, recall, accuracy, f1). -- This python routine also dumps out a large amount of images for convolution weights, and an image for {image_name}_output.png showing the classification.