A Study of Face Obfuscation in ImageNet
Code for the paper:
A Study of Face Obfuscation in ImageNet
Kaiyu Yang, Jacqueline Yau, Li Fei-Fei, Jia Deng, and Olga Russakovsky
@article{yang2021imagenetfaces,
title={A Study of Face Obfuscation in ImageNet},
author={Yang, Kaiyu and Yau, Jacqueline and Fei-Fei, Li and Deng, Jia and Russakovsky, Olga},
journal={arXiv preprint arXiv:2103.06191},
year={2021}
}
Face Annotation
crowdsourcing/ui.html is the UI used for face annotation. It should be used as an HTML template in simple-amt. Please refer to the documentation of simple-amt for detail. The final face annotations are available for download.
Requirements
- Download and install Miniconda Python 3 (Anaconda should also work).
- Edit imagenet-face-obfuscation.yaml according to your system. For example, remove - cudatoolkit=11.0 if you don't have a GPU. Change the version of cudatoolkit if necessary. You could see the instructions for installing PyTorch for what CUDA version to put in imagenet-face-obfuscation.yaml.
- Install Python dependencies using conda:
conda env create -f imagenet-face-obfuscation.yaml && conda activate imagenet-face-obfuscation
. If you have troubles with the aforementioned two steps, you may manually install the packages in imagenet-face-obfuscation.yaml in whatever way that works for you.
Data
The face-blurred images are avaialble for download on the ImageNet website. They were generated by running python experiments/blurring.py
. Please refer to the source file of experiments/blurring.py for details. For original images, please use the official ILSVRC 2012 dataset.
Save the original images to data/train/
and data/val/
; save the blurred images to data/train_blurred/
and data/val_blurred/
. In each directory, each category should has a subdirectory, and images should be in the subdirectories. For example: data/val_blurred/n02119022/ILSVRC2012_val_00012978.jpg
.
Training and Validation
experiments/trainval.py is the script for training and validation. It is based on an example from PyTorch with only minor changes. Most command-line options in the original example still apply. Please refer to the original documentation for details.
For example, to train a ResNet18 on a single GPU.
python experiments/trainval.py -a resnet18 --learning-rate 0.1 --gpu 0
To train a ResNet50 on all GPUs on the current node:
python experiments/trainval.py -a resnet50 --learning-rate 0.1 --dist-url 'tcp://127.0.0.1:6666' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0
We add a few additional command-line options for training/evaluating on face-obfuscated images:
--exp-id EXPID
: "EXPID" is an arbitrary experiment identifier. Model checkpoints will be saved toEXPID_model_best.pth
andEXPID_model_latest.pth
. Validation results will be saved toEXPID_val_results.pickle
.--blur-train
: Use face-blurred images for training.--blur-val
: Use face-blurred images for validation.--overlay
: Use overlayed images for both training and validation. It cannot co-occur with--blur-train
or--blur-val
.
For example, to train and evaluate an AlexNet on face-blurred images:
python experiments/trainval.py -a alexnet --learning_rate 0.01 --gpu 0 --blur-train --blur-val --exp-id alexnet_blurred_train_blurred_val
To train a ResNet152 on face-blurred images but evalaute on original images:
python experiments/trainval.py -a resnet152 --learning-rate 0.1 --dist-url 'tcp://127.0.0.1:6667' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0 --blur-train --exp-id hello_my_experiment
Models pretrained on face-blurred images are available for download here.
Our validation results for all models are available here. Before the next step, please download these pickle files to eval_pickles/
. You could also run the training script to produce them by yourself.
Analyses
Please first make sure validation pickle files are in eval_pickles/
and face annotations are in data/face_annotations_ILSVRC.json
.
Faces in different supercategories
To produce Table 2 in the paper:
python analysis/supercategories.py
Supercategory #Categories #Images With faces (%)
----------------------- ------------- --------- ----------------
clothing.n.01 49 62471 58.9025
wheeled_vehicle.n.01 44 57055 35.2975
musical_instrument.n.01 26 33779 47.6361
bird.n.01 59 76536 1.68809
insect.n.01 27 35097 1.80642
Faces in different categories
To produce Figure 2 in the paper:
python analysis/num_images.py
Overall validation accuracy
To produce Table 3 in the paper:
python analysis/overall_accuracy.py
model top1 original top1 blurred top1 diff top5 original top5 blurred top5 diff
------------------ --------------- --------------- ----------- --------------- --------------- -----------
alexnet 56.043 +- 0.258 55.834 +- 0.108 0.209 78.835 +- 0.115 78.547 +- 0.071 0.288
squeezenet1_0 55.989 +- 0.179 55.323 +- 0.039 0.666 78.602 +- 0.172 78.061 +- 0.017 0.541
shufflenet_v2_x1_0 64.646 +- 0.178 64.001 +- 0.068 0.645 85.927 +- 0.024 85.458 +- 0.051 0.47
vgg11 68.905 +- 0.039 68.209 +- 0.128 0.695 88.682 +- 0.025 88.283 +- 0.046 0.399
vgg13 69.925 +- 0.058 69.271 +- 0.103 0.653 89.324 +- 0.064 88.928 +- 0.034 0.396
vgg16 71.657 +- 0.061 70.839 +- 0.047 0.818 90.456 +- 0.067 89.897 +- 0.108 0.559
vgg19 72.363 +- 0.023 71.538 +- 0.032 0.826 90.866 +- 0.053 90.289 +- 0.008 0.577
mobilenet_v2 65.378 +- 0.182 64.367 +- 0.203 1.011 86.651 +- 0.059 85.969 +- 0.060 0.682
densenet121 75.036 +- 0.055 74.244 +- 0.064 0.792 92.375 +- 0.031 91.958 +- 0.100 0.417
densenet201 76.984 +- 0.021 76.551 +- 0.044 0.433 93.480 +- 0.034 93.223 +- 0.068 0.257
resnet18 69.767 +- 0.171 69.012 +- 0.174 0.755 89.223 +- 0.024 88.738 +- 0.031 0.485
resnet34 73.083 +- 0.131 72.307 +- 0.351 0.776 91.289 +- 0.008 90.755 +- 0.130 0.534
resnet50 75.461 +- 0.198 75.003 +- 0.074 0.458 92.487 +- 0.015 92.360 +- 0.071 0.127
resnet101 77.254 +- 0.070 76.735 +- 0.092 0.519 93.591 +- 0.085 93.310 +- 0.052 0.281
resnet152 77.853 +- 0.117 77.279 +- 0.091 0.573 93.933 +- 0.038 93.674 +- 0.011 0.26
average 70.023 69.368 0.655 89.048 88.630 0.418
To produce Table B in the paper:
python analysis/overall_accuracy_overlay.py
model top1 original top1 overlayed top1 diff top5 original top5 overlayed top5 diff
------------------ --------------- ----------------- ----------- --------------- ----------------- -----------
alexnet 56.043 +- 0.258 55.474 +- 0.236 0.569 78.835 +- 0.115 78.172 +- 0.187 0.663
squeezenet1_0 55.989 +- 0.179 55.039 +- 0.221 0.95 78.602 +- 0.172 77.633 +- 0.108 0.969
shufflenet_v2_x1_0 64.646 +- 0.178 63.684 +- 0.033 0.962 85.927 +- 0.024 85.166 +- 0.167 0.761
vgg11 68.905 +- 0.039 67.834 +- 0.157 1.071 88.682 +- 0.025 87.880 +- 0.036 0.802
vgg13 69.925 +- 0.058 68.749 +- 0.015 1.175 89.324 +- 0.064 88.536 +- 0.062 0.788
vgg16 71.657 +- 0.061 70.568 +- 0.100 1.089 90.456 +- 0.067 89.573 +- 0.019 0.883
vgg19 72.363 +- 0.023 71.206 +- 0.152 1.158 90.866 +- 0.053 90.104 +- 0.050 0.762
mobilenet_v2 65.378 +- 0.182 64.335 +- 0.162 1.043 86.651 +- 0.059 85.728 +- 0.066 0.922
densenet121 75.036 +- 0.055 74.062 +- 0.048 0.974 92.375 +- 0.031 91.700 +- 0.025 0.675
densenet201 76.984 +- 0.021 76.056 +- 0.073 0.928 93.480 +- 0.034 92.868 +- 0.064 0.612
resnet18 69.767 +- 0.171 68.938 +- 0.069 0.829 89.223 +- 0.024 88.665 +- 0.110 0.557
resnet34 73.083 +- 0.131 72.369 +- 0.099 0.714 91.289 +- 0.008 90.699 +- 0.020 0.589
resnet50 75.461 +- 0.198 74.916 +- 0.007 0.545 92.487 +- 0.015 92.154 +- 0.027 0.333
resnet101 77.254 +- 0.070 76.677 +- 0.102 0.577 93.591 +- 0.085 93.114 +- 0.077 0.476
resnet152 77.853 +- 0.117 76.978 +- 0.149 0.875 93.933 +- 0.038 93.342 +- 0.246 0.592
average 70.023 69.126 0.897 89.048 88.356 0.692
Category-wise accuracies
To produce Table 4 in the paper:
python analysis/categorywise_accuracies.py
Category top1 original top1 blurred top1 diff top5 original top5 blurred top5 diff AP original AP blurred AP diff
----------------------- --------------- --------------- ----------- --------------- --------------- ----------- --------------- --------------- ---------
eskimo_dog.n.01 50.800 +- 1.105 37.956 +- 0.412 12.844 95.467 +- 0.377 95.156 +- 0.166 0.311 19.378 +- 0.765 19.908 +- 0.481 -0.529
siberian_husky.n.01 46.267 +- 1.792 63.200 +- 0.762 -16.933 96.978 +- 0.440 97.244 +- 0.251 -0.267 29.198 +- 0.283 29.616 +- 0.485 -0.418
projectile.n.01 35.556 +- 0.880 21.733 +- 0.998 13.822 86.178 +- 0.412 85.467 +- 0.377 0.711 23.098 +- 0.365 22.537 +- 0.510 0.561
missile.n.01 31.556 +- 0.708 45.822 +- 0.817 -14.267 81.511 +- 0.725 81.822 +- 0.382 -0.311 20.404 +- 0.264 21.120 +- 0.633 -0.716
tub.n.02 35.511 +- 1.462 27.867 +- 0.576 7.644 79.422 +- 0.600 75.644 +- 0.453 3.778 19.853 +- 0.430 18.778 +- 0.231 1.075
bathtub.n.01 35.422 +- 0.988 42.533 +- 0.377 -7.111 78.933 +- 0.327 80.800 +- 1.236 -1.867 27.378 +- 0.757 25.079 +- 0.584 2.299
american_chameleon.n.01 62.978 +- 0.350 54.711 +- 1.214 8.267 96.978 +- 0.491 96.578 +- 0.453 0.4 39.963 +- 0.184 39.292 +- 0.525 0.671
green_lizard.n.01 42.000 +- 0.566 45.556 +- 1.238 -3.556 91.289 +- 0.274 89.689 +- 0.166 1.6 22.615 +- 0.775 22.407 +- 0.095 0.208
Credits
- The code for training and evaluating models on ILSVRC is based on an official PyTorch example.
- The code is formatted using .