The backgrounds challenge is a public dataset challenge for creating more background-robust models. This repository contains test datasets of ImageNet-9 (IN-9) with different amounts of background and foreground signal, which you can use to measure the extent to which your models rely on image backgrounds. These are described further in the paper: "Noise or Signal: The Role of Image Backgrounds in Object Recognition" (preprint, blog).
Deep computer vision models rely on both foreground objects and image backgrounds. Even when the correct foreground object is present, such models often make incorrect predictions when the image background is changed, and they are especially vulnerable to adversarially chosen backgrounds. For example, the the official pre-trained PyTorch ResNet-50 has an accuracy of 22% when evaluated against adversarial backgrounds on ImageNet-9 (for reference, a model that always predicts "dog" has an accuracy of 11%).
Thus, the goal of this challenge is to understand how background-robust models can be. Specifically, we assess models by their accuracy on images containing foregrounds superimposed on backgrounds which are adversarially chosen from the test set. We encourage researchers to use this challenge to benchmark progress on background-robustness, which can be important for determining models' out of distribution performance. We will maintain a leaderboard of top submissions.
| Model | Reference | Challenge Accuracy |
Clean Accuracy (on IN-9) |
Download Link |
|---|---|---|---|---|
| ResNet-50 | (initial entry) | 22.3% | 95.6% | Official Pytorch Model |
| ResNet-50 (IN-9L) | (initial entry) | 12.0% | 96.4% | Download |
To evaluate your model against adversarial backgrounds, you will need to do the following:
- Download and unzip the datasets included in the release.
- Run
python challenge_eval.py --checkpoint '/PATH/TO/CHECKPOINT' --data-path '/PATH/TO/DATA'.
The model checkpoint that the script takes as input must be one of the following.
- A 1000-class ImageNet classifier.
- A 9-class IN-9 classifier.
See python challenge_eval.py -h for how to toggle between the two.
Note: evaluation requires PyTorch to be installed with CUDA support.
We invite any interested researchers to submit models and results by submitting a pull request with your model checkpoint included. The most successful models will be listed in the leaderboard above. We have already included baseline pre-trained models for reference.
ImageNet-9 and its variations can be useful for measuring the impact of backgrounds on model decision making—see the paper for more details. You can test your own models on IN-9 and its variations as follows.
- Download and unzip the datasets included in the release.
- Run, for example,
python in9_eval.py --eval-dataset 'mixed_same' --checkpoint '/PATH/TO/CHECKPOINT' --data-path '/PATH/TO/DATA'. You can replacemixed_samewith whichever variation of IN-9 you are interested in.
Just like in the challenge, the input can either be a 1000-class ImageNet model or a 9-class IN-9 model.
There is no leaderboard or challenge for these datasets, but we encourage researchers to use these datasets to measure the role of image background in their models' decision making. Furthermore, we include a table of results for common pre-trained models and various models discussed in the paper.
| Model | Original | Mixed-Same | Mixed-Rand | BG-Gap |
|---|---|---|---|---|
| VGG16-BN | 94.3% | 83.6% | 73.4% | 10.2% |
| ResNet-50 | 95.6% | 86.2% | 78.9% | 7.3% |
| ResNet-152 | 96.7% | 89.3% | 83.5% | 5.8% |
| ResNet-50 (IN-9L) | 96.4% | 89.8% | 75.6% | 14.2% |
| ResNet-50 (IN-9/Mixed-Rand) | 73.3% | 71.5% | 71.3% | 0.2% |
The BG-Gap, or the difference between Mixed-Same and Mixed-Rand, measures the impact of background correlations in the presence of correct-labeled foregrounds.
Updated June 24, 2020: We are releasing all training data that we used to train models described in the paper. The download links are as follows: IN-9L, Mixed-Next, Mixed-Rand, Mixed-Same, No-FG, Only-BG-B, Only-BG-T, Only-FG, Original.
Each downloadable dataset contains both training data and validation data generated in the same way as the training data (that is, with no manual cleaning); this validation data can be safely ignored. The test data in the release should be used instead.
If you find these datasets useful in your research, please consider citing:
@article{xiao2020noise,
title={Noise or Signal: The Role of Image Backgrounds in Object Recognition},
author={Kai Xiao and Logan Engstrom and Andrew Ilyas and Aleksander Madry},
journal={ArXiv preprint arXiv:2006.09994},
year={2020}
}