tasks-rule

A study of the role of tasks in shaping visual processing

Tasks studied:

• Task 1 : Single-class classification task link.
  • Images are from MS-COCO.
  • There are 50,000 images in the training set and 10,000 images in the test set.
  • Pickle file organization structure = [training set, test_set], training set = [image id list, label list], test set = [image id list, label list].
  • image names: "./train2014/COCO_train2014_%012d.jpg" % (image_id)
• Task 2: Ten-class classification task
  • Images are from ImageNet.
  • The dataset consists of 60,000 images and ten categories. Each category contains 6,000 images.
  • The compressed file 10-class-classification-dataset.rar is in /data4/chenhaoran on img16.
  • There are ten folders in the compressed file and the name of each folder is the category name.
• Task 3: 20-class-multilabel classification task link
  • Images are from MS-COCO.
  • The dataset consists of 60, 000 images and twenty categories.
  • Pickle file organization structure = [dataset, label_description], dataset = [image id list, labels list], label description = {label: category name}.
  • image names: "./train2014/COCO_train2014_%012d.jpg" % (image_id) or "./val2014/COCO_val2014_%012d.jpg" % (image_id).
• Task 4: Ten-class classification task: inanimate
  • Images are from ImageNet.
  • The dataset consists of 13,000 images and ten categories. Each category contains 1,300 images.
  • The folder task4 is in /data4/chenhaoran on img16.
  • There are ten folders in the primary folder and the name of each folder is the category name.
  • The pickle file for task4 is also in /data4/chenhaoran on img16.
• Task 5: Ten-class classification task: animate
  • Images are from ImageNet.
  • The dataset consists of 15,600 images and ten categories.
  • The folder task5 is in /data4/chenhaoran on img16.
  • There are ten folders in the primary folder and the name of each folder is the category name.
  • The pickle file for task5 is also in /data4/chenhaoran on img16.

fMRI data:

The BOLD5000 dataset: fMRI responses of 4 subjects to 5254 images taken from COCO (2000 images), ImageNet (1916 images) and scene images (1000).

preprocessed images and fMRI data are now available in format imagefilename.pt : (image_array,fMRI_array), the fMRI array is the concatenation of the visual cortex voxels for subjects 1 to 3