search

grispeut / Feature-Alignment

object detection, adversarial robustness, ICIP2021
17 stars 1 forks source link

readme

Using Feature Alignment Can Improve Clean Average Precision and Adversarial Robustness in Object Detection

Paper:https://arxiv.org/abs/2012.04382

This repository contains two parts: one is experiments on YOLO-V3 in folder yolov3-adv, the other is experiments on FASTER-RCNN-FPN in folder faster-rcnn-fpn-adv.

Clone Code

git clone https://github.com/grispeut/Feature-Alignment.git

YOLO-V3

Follow the README in yolov3-adv to reproduce our experiments.

On Pascal VOC

model performance clean AP PGD-1 PGD-3 PGD-5 PGD-10 advAP acAP
STD 0.780 0.258 0.051 0.024 0.010 0.086 0.433
model performance clean AP PGD-1 PGD-3 PGD-5 PGD-10 advAP acAP
AT 0.724 0.301 0.187 0.167 0.148 0.201 0.463
TOD 0.702 0.298 0.195 0.177 0.163 0.208 0.455
KDFA 0.739 0.327 0.200 0.179 0.162 0.217 0.478
SSFA 0.730 0.297 0.175 0.152 0.133 0.189 0.460
FA 0.745 0.329 0.200 0.178 0.161 0.217 0.481
model performance clean AP PGD-1 PGD-3 PGD-5 PGD-10 advAP acAP
AT 0.665 0.277 0.209 0.196 0.186 0.217 0.441
TOD 0.659 0.285 0.219 0.208 0.199 0.228 0.444
KDFA 0.702 0.315 0.229 0.213 0.201 0.240 0.471
SSFA 0.693 0.293 0.210 0.196 0.185 0.221 0.457
FA 0.712 0.319 0.231 0.213 0.202 0.241 0.477
model performance clean AP PGD-1 PGD-3 PGD-5 PGD-10 advAP acAP
AT 0.604 0.251 0.194 0.183 0.176 0.201 0.403
TOD 0.581 0.249 0.200 0.190 0.183 0.206 0.394
KDFA 0.672 0.301 0.231 0.219 0.210 0.240 0.456
SSFA 0.653 0.281 0.214 0.203 0.194 0.223 0.438
FA 0.677 0.305 0.231 0.218 0.209 0.241 0.459

On MS-COCO

model performance clean AP PGD-1 PGD-3 PGD-5 PGD-10 advAP acAP
STD 0.545 0.189 0.041 0.022 0.009 0.065 0.305
model performance clean AP PGD-1 PGD-3 PGD-5 PGD-10 advAP acAP
AT 0.490 0.187 0.077 0.056 0.037 0.089 0.290
TOD 0.488 0.191 0.073 0.052 0.035 0.088 0.288
KDFA 0.506 0.209 0.098 0.075 0.059 0.110 0.308
SSFA 0.499 0.202 0.089 0.068 0.051 0.103 0.301
FA 0.510 0.213 0.109 0.087 0.070 0.120 0.315

FASTER-RCNN-FPN

Follow the README in faster-rcnn-fpn-adv to reproduce our experiments.

On Pascal VOC

model performance clean AP PGD-1 PGD-3 PGD-5 PGD-10 advAP acAP
STD 0.835 0.265 0.009 0.002 0.002 0.070 0.453
model performance clean AP PGD-1 PGD-3 PGD-5 PGD-10 advAP acAP
AT 0.834 0.189 0.065 0.035 0.007 0.074 0.454
TOD 0.833 0.309 0.073 0.037 0.006 0.106 0.470
KDFA 0.833 0.354 0.103 0.059 0.015 0.133 0.483
SSFA 0.833 0.345 0.078 0.053 0.015 0.123 0.478
FA 0.834 0.480 0.109 0.064 0.010 0.166 0.500
model performance clean AP PGD-1 PGD-3 PGD-5 PGD-10 advAP acAP
AT 0.824 0.548 0.209 0.154 0.031 0.236 0.530
TOD 0.828 0.579 0.224 0.161 0.024 0.247 0.538
KDFA 0.828 0.645 0.241 0.178 0.029 0.273 0.551
SSFA 0.829 0.623 0.235 0.178 0.033 0.267 0.548
FA 0.832 0.664 0.256 0.189 0.033 0.286 0.559
model performance clean AP PGD-1 PGD-3 PGD-5 PGD-10 advAP acAP
AT 0.822 0.281 0.153 0.103 0.009 0.137 0.480
TOD 0.821 0.367 0.184 0.131 0.010 0.173 0.497
KDFA 0.823 0.452 0.232 0.164 0.018 0.217 0.520
SSFA 0.824 0.404 0.202 0.141 0.012 0.190 0.507
FA 0.825 0.545 0.265 0.191 0.039 0.260 0.543