Open jsherrah opened 10 years ago
I've created a bunch of issues rather than using the todo list spreadsheet, which was clunky.
It turns out I wasn't getting emails for issues because I'm the repo owner. I've created a new organisation, made us both members of the organisation and added the repo to the organisation. Like the name?
Here are the params I got from grid search on all MSRC data, random forest:
Done. Grid search gave these parameters: max_features : 25 n_estimators : 50 max_depth : 100 min_samples_leaf : 5
I wrote a bash script that runs the classifier over all the msrc images, creates the output labelled images and a csv with the filenames that you need:
./classifyAllImages.sh msrcFull_randForest_grid.pkl results/imagesClassified ~/data/sceneLabelling/MSRC_ObjCategImageDatabase_v2/Images/*.bmp
To run MRF with multiple K parameters:
for K in .1 .2 .4 .8 1; do echo $K; ./sceneLabelSuperPixels.py msrcFull_randForest_grid.pkl 3_7_s.bmp --outfile testmrf.png --K=$K --adjFn=msrcFull_train_adj.pkl --nbrPotentialMethod=adjacencyAndDegreeSensitive; display testmrf.png; done
Read me feedback
I had permission issues.
Running as suo failed. Was necessary to chown the following files to the current user:
/
/
Big success! Created and pushed a Vagrantfile with old and new syntax for config to repo. VM didn't auto-provision on first load, might be best to recommend
vagrant up --provision
for first use in the ReadMe. Both vagrant ssh and default synched folder worked without issue.
Thanks for this solution, looks very promising. How to use it with super pixels (have average color, stddev, centroid of super pixel ) or custom data? Is there any tutorial for preparing Input data?
Thanks for the question, yes the code does work with superpixels, and computes feature summaries over the superpixels. Unfortunately there is no tutorial or comprehensive documentation at this stage, you'll have to look at what examples there are with the code.
Here's a forum for back-and-forth discussions (blog)