A pytorch implementation for paper "Iterative Quantization: A Procrustean Approach to Learning Binary Codes for Large-scale Image Retrieval" TPAMI-2013

REQUIREMENTS

pip install -r requirements.txt

1. pytorch >= 1.0
2. loguru

DATASETS

1. cifar10-gist.mat password: umb6
2. cifar-10_alexnet.t password: f1b7
3. nus-wide-tc21_alexnet.t password: vfeu
4. imagenet-tc100_alexnet.t password: 6w5i

USAGE

usage: run.py [-h] [--dataset DATASET] [--root ROOT]
              [--code-length CODE_LENGTH] [--max-iter MAX_ITER] [--topk TOPK]
              [--gpu GPU]

ITQ_PyTorch

optional arguments:
  -h, --help            show this help message and exit
  --dataset DATASET     Dataset name.
  --root ROOT           Path of dataset
  --code-length CODE_LENGTH
                        Binary hash code length.(default:
                        8,16,24,32,48,64,96,128)
  --max-iter MAX_ITER   Number of iterations.(default: 3)
  --topk TOPK           Calculate map of top k.(default: ALL)
  --gpu GPU             Using gpu.(default: False)

EXPERIMENTS

cifar10-gist dataset. Gist features, 1000 query images, 5000 training images, MAP@ALL.

cifar-10-alexnet dataset. Alexnet features, 1000 query images, 5000 training images, MAP@ALL.

nus-wide-tc21-alexnet dataset. Alexnet features, top 21 classes, 2100 query images, 10500 training images, MAP@5000.

imagenet-tc100-alexnet dataset. Alexnet features, top 100 classes, 5000 query images, 10000 training images, MAP@1000.