Skip to content

Latest commit

 

History

History
123 lines (100 loc) · 5.45 KB

README.md

File metadata and controls

123 lines (100 loc) · 5.45 KB
Raw

Base datasets

For a few datasets that FsDet natively supports, the datasets are assumed to exist in a directory called "datasets/", under the directory where you launch the program. They need to have the following directory structure:

Expected dataset structure for Pascal VOC:

VOC20{07,12}/
  Annotations/
  ImageSets/
  JPEGImages/

Expected dataset structure for COCO:

coco/
  annotations/
    instances_{train,val}2014.json
  {train,val}2014/
    # image files that are mentioned in the corresponding json

Expected dataset structure for LVIS:

coco/
  {train,val}2017/
lvis/
  lvis_v0.5_{train,val}.json
  lvis_v0.5_train_{freq,common,rare}.json

LVIS uses the same images and annotation format as COCO. You can use split_lvis_annotation.py to split lvis_v0.5_train.json into lvis_v0.5_train_{freq,common,rare}.json.

Install lvis-api by:

pip install git+https://github.com/lvis-dataset/lvis-api.git

Few-shot datasets

For each dataset, we additionally create few-shot versions by sampling shots for each novel category. For better comparisons, we sample multiple groups of training shots in addition to the ones provided in previous works. We include the sampling scripts we used for better reproducibility and extensibility. The few-shot dataset files can be found here. They should have the following directory structure:

Pascal VOC:

vocsplit/
  box_{1,2,3,5,10}shot_{category}_train.txt
  seed{1-29}/
    # shots

Each file contains the images for K shots for a specific category. There may be more instances in the images than K; in these cases, we randomly sample K instances.

The shots in the vocsplit directory are the same shots used by previous works. We additionally sample 29 more groups of shots for a total of 30 groups, which can be generated by using prepare_voc_few_shot.py.

See prepare_voc_few_shot.py for generating the seeds yourself.

In the config files, you have to specify the datasets to use. See our provided config files for more usage details. Below, we list out the names of the datasets and the data they correspond to:

voc_20{07,12}_trainval_{base,all}{1,2,3}        # Train/val datasets with base categories or all
                                                  categories for splits 1, 2, and 3.
voc_2007_trainval_all{1,2,3}_{1,2,3,5,10}shot   # Balanced subsets for splits 1, 2, and 3 containing
                                                  1, 2, 3, 5, or 10 shots for each category. You only
                                                  need to specify 2007, as it will load in both 2007
                                                  and 2012 automatically.
voc_2007_trainval_novel{1,2,3}_{1,2,3,5,10}shot # Same as previous datasets, but only contains data
                                                  of novel categories.
voc_2007_test_{base,novel,all}{1,2,3}           # Test datasets with base categories, novel categories,
                                                  or all categories for splits 1, 2, and 3.

COCO:

cocosplit/
  datasplit/
    trainvalno5k.json
    5k.json
  full_box_{1,2,3,5,10,30}shot_{category}_trainval.json
  seed{1-9}/
    # shots

All but 5k images from the train/val sets are used for training, while 5k images from the val set are used as the test set. trainvalno5k.json denotes the training set and 5k.json is the test set.

The sampling procedure is the same as for Pascal VOC, except we sample exactly K instances for each category. For COCO, we use 10 groups.

See prepare_coco_few_shot.py for generating the seeds yourself.

Dataset names for config files:

coco_trainval_{base,all}                        # Train/val datasets with base categories or all
                                                  categories.
coco_trainval_all_{1,2,3,5,10,30}shot           # Balanced subsets containing 1, 2, 3, 5, 10, or 30
                                                  shots for each category.
coco_trainval_novel_{1,2,3,5,10,30}shot         # Same as previous datasets, but only contains data
                                                  of novel categories.
coco_test_{base,novel,all}                      # Test datasets with base categories, novel categories,
                                                  or all categories.

LVIS:

lvissplit/
  lvis_shots.json

We treat the frequent and common categories as the base categories and the rare categories as the novel categories.

We sample up to 10 instances for each category to build a balanced subset for the few-shot fine-tuning stage. We include all shots in a single COCO-style annotation file.

See prepare_lvis_few_shot.py for generating the seeds yourself.

Dataset names for config files:

lvis_v0.5_train_{freq,common}                   # Train datasets with freq categories or common
                                                  categories. These are used as the base datasets.
lvis_v0.5_train_rare_novel                      # Train datasets with rare categories.
lvis_v0.5_train_shots                           # Balanced subset containing up to 10 shots for
                                                  each category.
lvis_v0.5_val                                   # Validation set with all categories.
lvis_v0.5_val_novel                             # Validation set with only novel categories.