Zero Shot Detection Dataset

This repository contains the datasets used in Zero Shot Detection by Pengkai Zhu, Hanxiao Wang, Tolga Bolukbasi and Venkatesh Saligrama. The pdf version is available here.

@article{zhu2018zero,
  title={Zero-Shot Detection},
  author={Zhu, Pengkai and Wang, Hanxiao and Bolukbasi, Tolga and Saligrama, Venkatesh},
  journal={arXiv preprint arXiv:1803.07113},
  year={2018}
}

Zero Shot Detection

We propose a new Zero-Shot Detection (ZSD) problem, referring to the task of detecting classes with zero training data.

To solve the ZSD problem, we propose a novel zero-shot method based on training an end-to-end model that fuses semantic attribute prediction with visual features to propose object bounding boxes for seen and unseen classes. Our method retains the efficiency and effectiveness of YOLO for objects seen during training, while improving its performance for novel and unseen objects.

Seen/Unseen Split

The scripts can download Pascal VOC or MSCOCO and split it into four parts as in the paper:

• Train: seen in train
• Test-Seen: seen in val/test
• Test-Unseen: unseen in train&val&test
• Test-Mix: both seen & unseen in val/test

The dataset is split based on assigned seen categories names. We provide the splits we used in the paper in seen_names subfolder.

Attributes

The attributes for Pascal VOC will be downloaded and extracted automatically when running get_voc_zsd_dataset.sh. We also provide the attributes we use in the paper in the attributes subfolder:

• coco_w2v.txt: w2v attributes for coco categories
• coco_w2v_voc.txt: projected w2v attributes for coco categories (mirroring VOC attributes similarity)
• voc.txt: labelled attributes (from aP&Y) for VOC categories
• voc_w2v.txt: w2v attributes for VOC categories

How to use

• Preliminary: numpy

Setup Pascal VOC:

bash get_voc_zsd_dataset.sh $zsd-data-dir  # $zsd-data-dir: directory for saving pascal ZSD dataset

The dataset will be downloaded to $zsd-data-dir and the split sets will be saved in 1010split subfolder by default. If you already downloaded the dataset or would like to try some other splits, just run:

python zsd_split.py --dataset voc --data_dir $zsd-data-dir --name_file voc.names \
--seen_name_file seen_names/voc/${choose another split} \
--save_dir ${split save name} \

Setup MSCOCO

bash get_coco_zsd_dataset.sh $zsd-data-dir  # $zsd-data-dir: directory for saving coco ZSD dataset