VisionLearningGroup / SANE

Pytorch implementation of our ECCV paper

Why do These Match? Explaining the Behavior of Image Similarity Models (ECCV 2020)

This repository implements:

Bryan A. Plummer*, Mariya I. Vasileva*, Vitali Petsiuk, Kate Saenko, David Forsyth.

Why do These Match? Explaining the Behavior of Image Similarity Models. ECCV, 2020.

Environment

This code was tested with Python 3.6 and Pytorch 1.4.

Preparation

Download data and unzip it in ./data

• Polyvore Outfits
• Animals with Attributes 2

Saliency Map Test

To reproduce results from Table 1 in our paper you can use the saliency.py script in tools, e.g.,

python ./tools/saliency.py --fixed_ref --dataset polyvore_outfits --method rise

Training Attribute Classifier

1. Cache saliency maps (used in Eq.2 of our paper)

python ./tools/saliency.py --fixed_ref --dataset polyvore_oufits --method rise --split train

2. Train the model

python ./tools/train_attribute_classifier.py --fixed_ref --dataset polyvore_outfits --method rise

Citation

If you find our code useful please consider citing:

@InProceedings{plummerSimilarityExplanations2020,
     author={Bryan A. Plummer and Mariya I. Vasileva and Vitali Petsiuk and Kate Saenko and David Forsyth},
     title={Why do These Match? Explaining the Behavior of Image Similarity Models},
     booktitle = {The European Conference on Computer Vision (ECCV)},
     year = {2020}
}