Deep-RTC [project page]

This repository contains the source code accompanying our ECCV 2020 paper.

Solving Long-tailed Recognition with Deep Realistic Taxonomic Classifier
Tz-Ying Wu, Pedro Morgado, Pei Wang, Chih-Hui Ho, Nuno Vasconcelos

@inproceedings{Wu20DeepRTC,
	title={Solving Long-tailed Recognition with Deep Realistic Taxonomic Classifier},
	author={Tz-Ying Wu and Pedro Morgado and Pei Wang and Chih-Hui Ho and Nuno Vasconcelos},
	booktitle={European Conference on Computer Vision (ECCV)},
	year={2020}
}

Dependencies

• Python (3.5.6)
• PyTorch (1.2.0)
• torchvision (0.4.0)
• NumPy (1.15.2)
• Pillow (5.2.0)
• PyYaml (5.1.2)
• tensorboardX (1.8)

Data preparation

• CIFAR100 [Raw images] [Long-tail version]
• AWA2 [Raw images]
• ImageNet [Raw images] [Long-tail version]
• iNaturalist [Raw images]

These datasets can be downloaded from the above links. Please organize the images in the hierarchical folders that represent the dataset hierarchy, and put the root folder under prepro/raw. For example,

prepro/raw/imagenet
--abstraction
----bubble
------ILSVRC2012_val_00014026.JPEG
------ILSVRC2012_val_00000697.JPEG
...
--physical_entity
----object
...

While CIFAR100 and iNaturalist have released taxonomies, we built the tree-type taxonomy of AWA2 and ImageNet with WordNet. All the taxonomies are provided in prepro/data/{dataset}/tree.npy, and the data splits are provided in prepro/splits/{dataset}/{split}.json. Please refer to prepro/README.md for more details. After the raw images are managed hierarchically, run

$ ./prepare_data.sh {dataset}

where {dataset}=awa2/cifar100/imagenet/inaturalist. This will automatically generate the data lists for all splits, and build the codeword matrices needed for training Deep-RTC. Note that our codes can be applied to other datasets once they are organized hierarchically.

Training and evaluation

To train and evaluate Deep-RTC, run

$ export PYTHONPATH=${PWD}/prepro:${PYTHONPATH}
$ ./run.sh {dataset}

where {dataset}=awa2/cifar100/imagenet/inaturalist. Our pretrained models can be downloaded here.