This repository contains implementation of the following paper:

Deep Metric Learning to Rank
Fatih Cakir*, Kun He*, Xide Xia, Brian Kulis, and Stan Sclaroff (*equal contribution)
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019

FastAPLoss from pytorch-metric-learning

• Matlab: see matlab/README.md
• PyTorch: see pytorch/README.md

• Stanford Online Products
  • Can be downloaded here
• In-Shop Clothes Retrieval
  • Can be downloaded here
• PKU VehicleID
  • Please request the dataset from the authors here

• We provide trained MatConvNet models and experimental logs for the results in the paper. These models were used to achieve the results in the tables.
• The logs also include parameters settings that enable one to re-train a model if desired. It also includes evaluation results with model checkpoints at certain epochs.

  • Table 1: Stanford Online Products

    • FastAP, ResNet-18, M=256, Dim=512: [model @ epoch 20, log]
    • FastAP, ResNet-50, M=96, Dim=128: [model @ epoch 30, log]
    • FastAP, ResNet-50, M=96, Dim=512: [model @ epoch 28, log]
    • FastAP, ResNet-50, M=256, Dim=512: [model @ epoch 12, log]

  • Table 2: In-Shop Clothes

    • FastAP, ResNet-18, M=256, Dim=512: [model @ epoch 50, log]
    • FastAP, ResNet-50, M=96, Dim=512: [model @ epoch 40, log]
    • FastAP, ResNet-50, M=256, Dim=512: [model @ epoch 35, log]

  • Table 3: PKU VehicleID

    • FastAP, ResNet-18, M=256, Dim=512: [model @ epoch 50, log]
    • FastAP, ResNet-50, M=96, Dim=512: [model @ epoch 40, log]
    • FastAP, ResNet-50, M=256, dim=512: [model @ epoch 30, log]

    (M=mini-batch size)

• PyTorch code is a direct port from our MATLAB implementation. We haven't tried reproducing the paper results with our PyTorch code. For reproducibility use the MATLAB version.
• Note that the mini-batch sampling strategy must also be used alongside the FastAP loss for good results.

For questions and comments, feel free to contact: kunhe@fb.com or fcakirs@gmail.com

MIT