occlusion-project

  • pbs_script/: Scripts for running code on cluster with PBS.
  • slurm_script/: Scripts for running code on cluster with slurm.
  • data/: Data folder, including flickr, imagenet and PASCAL 3D+. Included in .gitignore.
  • result/: Result of the experiments. Included in .gitignore.
  • src/: Source code for current project. (A.ipynb → B.py: using A to generate B, and running B on cluster with PBS / slurm scripts.)
    • Settings
      • settings.py
      • constant.py
    • Prepare datasets
      • dataset.ipynb → dataset.py: Generate training dataset and test dataset. Basically first 3 cells are used.
    • Finetune
      • finetune.ipynb → finetune.py: Create network files, finetune networks and visualize middle results.
    • Test
      • test_lmdb.ipynb → test.py: Generate test_{}_{}.prototxt for testing, and test model by reading images from lmdb.
      • test.ipynb → test.py: Test model by reading images directly from the disk. [Deprecated]
      • data_utility.ipynb: Some utilities to show and test selected images in datasets.
    • Analyse
      • img2vec.ipynb → img2vec.py: Extract feature vectors of images.
      • visualize_finetune: Plot training loss, training accuracy and validation accuracy during finetuning.
      • visualize_accuracy: Plot accuracy curves according to different occlusion levels.
      • visualize_img2vec: Plot feature space, and accuracy improvement for each class.
      • visualize_aperture: Plot results of aperture occlusion. [Deprecated]
      • vec2accuracy_divide: Test results with divided test datasets to get variation. [Deprecated]
      • heat_map.ipynb: Plot confidence heat maps of given images.
    • Miscellaneous
      • estimate_time.ipynb
  • legacy/: Legacy code related to visual concept, not used by current project.
    • legacy/src_visual_concept/: Use visual concept to generate occluded images.
    • legacy/src_cl/: Chen Liu's code of Picasso project, modified by Hao Wang.
    • legacy/src_orignal/: Chen Liu's original code of Picasso project.