VietPhong/kitti-yolo11n-robustness-benchmark

KITTI YOLO11n Robustness & Adversarial Benchmark Suite

This dataset contains 649,425 benchmark samples evaluating the perception robustness of YOLO11n (Ultralytics YOLOv11 nano in original FP32 precision) on the official KITTI Object Detection train set (3,711 images) under 35 attack & corruption techniques across 5 severity levels.

?? Benchmark Leaderboard (mAP@0.5 Drop on YOLO11n)

• Clean Baseline AP50: 0.3555
• Evaluation Model: YOLO11n (Original weights: yolo11n.pt, FP32 precision, size: 640x640)
• Target Classes: Car, Pedestrian, Cyclist

Attack Name	Group	Sev 1 AP (Drop)	Sev 2 AP (Drop)	Sev 3 AP (Drop)	Sev 4 AP (Drop)	Sev 5 AP (Drop)	Mean $\Delta$mAP
brightness	A	0.350 (-1.5%)	0.344 (-3.2%)	0.338 (-5.0%)	0.335 (-5.8%)	0.332 (-6.7%)	-0.0158
contrast	A	0.316 (-11.1%)	0.298 (-16.1%)	0.274 (-23.1%)	0.204 (-42.6%)	0.080 (-77.5%)	-0.1211
cw_l2	D	0.335 (-5.7%)	0.255 (-28.3%)	0.193 (-45.6%)	0.162 (-54.4%)	0.153 (-57.0%)	-0.1358
defocus_blur	A	0.347 (-2.3%)	0.328 (-7.7%)	0.286 (-19.4%)	0.259 (-27.1%)	0.220 (-38.1%)	-0.0673
depth_fog	B	0.291 (-18.1%)	0.213 (-40.1%)	0.036 (-89.9%)	0.000 (-99.9%)	0.000 (-99.9%)	-0.2474
depth_rain	B	0.352 (-1.0%)	0.352 (-1.0%)	0.348 (-2.0%)	0.345 (-3.1%)	0.338 (-5.0%)	-0.0086
depth_snow	B	0.351 (-1.3%)	0.344 (-3.2%)	0.326 (-8.2%)	0.314 (-11.8%)	0.210 (-41.0%)	-0.0466
dpatch	E	0.343 (-3.5%)	0.334 (-6.0%)	0.325 (-8.7%)	0.307 (-13.7%)	N/A	-0.0283
elastic_transform	A	0.352 (-0.9%)	0.352 (-1.0%)	0.341 (-4.1%)	0.341 (-4.2%)	0.325 (-8.7%)	-0.0136
fgsm	D	0.226 (-36.3%)	0.185 (-48.0%)	0.158 (-55.5%)	0.138 (-61.2%)	0.097 (-72.6%)	-0.1946
fog	A	0.309 (-13.0%)	0.300 (-15.6%)	0.287 (-19.2%)	0.292 (-17.7%)	0.273 (-23.2%)	-0.0630
frame_freeze	C	0.352 (-1.1%)	0.322 (-9.6%)	0.287 (-19.4%)	0.238 (-33.0%)	0.208 (-41.6%)	-0.0745
frost	A	0.312 (-12.4%)	0.264 (-25.8%)	0.228 (-36.0%)	0.211 (-40.8%)	0.186 (-47.5%)	-0.1155
gaussian_blur	A	0.354 (-0.5%)	0.335 (-5.8%)	0.304 (-14.4%)	0.273 (-23.1%)	0.207 (-41.8%)	-0.0609
gaussian_noise	A	0.330 (-7.1%)	0.317 (-10.7%)	0.284 (-20.0%)	0.244 (-31.2%)	0.172 (-51.8%)	-0.0859
glass_blur	A	0.307 (-13.5%)	0.308 (-13.3%)	0.231 (-35.0%)	0.207 (-41.7%)	0.153 (-56.8%)	-0.1140
impulse_noise	A	0.278 (-21.8%)	0.228 (-36.0%)	0.178 (-50.0%)	0.094 (-73.5%)	0.021 (-94.1%)	-0.1957
jpeg_compression	A	0.339 (-4.6%)	0.317 (-10.8%)	0.318 (-10.6%)	0.305 (-14.1%)	0.284 (-20.1%)	-0.0428
mi_fgsm	D	0.192 (-46.0%)	0.131 (-63.2%)	0.096 (-73.1%)	0.070 (-80.2%)	0.045 (-87.3%)	-0.2488
motion_blur	A	0.349 (-1.8%)	0.320 (-10.1%)	0.277 (-22.1%)	0.179 (-49.6%)	0.101 (-71.6%)	-0.1104
object_occlusion	C	0.322 (-9.3%)	0.186 (-47.8%)	0.004 (-98.8%)	0.000 (-100.0%)	0.000 (-100.0%)	-0.2531
pgd	D	0.149 (-58.0%)	0.096 (-73.2%)	0.060 (-83.2%)	0.039 (-88.9%)	0.024 (-93.3%)	-0.2820
pixelate	A	0.352 (-1.1%)	0.354 (-0.4%)	0.343 (-3.4%)	0.306 (-14.0%)	0.289 (-18.6%)	-0.0266
random_erasing	C	0.338 (-5.0%)	0.322 (-9.5%)	0.296 (-16.8%)	0.264 (-25.7%)	0.245 (-31.2%)	-0.0627
random_noise_linf	F	0.353 (-0.8%)	0.353 (-0.8%)	0.352 (-1.0%)	0.349 (-1.7%)	0.335 (-5.8%)	-0.0072
saturate	A	0.345 (-3.1%)	0.337 (-5.1%)	0.342 (-3.9%)	0.310 (-12.7%)	0.302 (-15.0%)	-0.0283
sensor_fault	C	0.352 (-0.9%)	0.326 (-8.3%)	0.306 (-14.0%)	0.280 (-21.2%)	0.211 (-40.6%)	-0.0604
shot_noise	A	0.317 (-10.8%)	0.288 (-18.8%)	0.241 (-32.3%)	0.150 (-57.9%)	0.074 (-79.3%)	-0.1416
snow	A	0.256 (-28.0%)	0.214 (-39.9%)	0.169 (-52.6%)	0.125 (-64.7%)	0.162 (-54.4%)	-0.1704
spatter	A	0.352 (-1.1%)	0.315 (-11.3%)	0.264 (-25.8%)	0.270 (-24.1%)	0.206 (-42.0%)	-0.0740
speckle_noise	A	0.343 (-3.4%)	0.335 (-5.7%)	0.294 (-17.2%)	0.269 (-24.2%)	0.238 (-33.0%)	-0.0594
square_attack	F	0.335 (-5.8%)	0.337 (-5.3%)	0.338 (-5.0%)	N/A	N/A	-0.0191
thys_patch	E	0.350 (-1.6%)	0.343 (-3.6%)	0.333 (-6.3%)	0.327 (-8.1%)	N/A	-0.0174
tog	D	0.148 (-58.4%)	0.101 (-71.7%)	0.066 (-81.5%)	0.035 (-90.1%)	0.024 (-93.4%)	-0.2809
zoom_blur	A	0.228 (-35.9%)	0.177 (-50.1%)	0.138 (-61.3%)	0.108 (-69.7%)	0.076 (-78.7%)	-0.2102

?? Dataset Structure

Each sample is stored in Apache Parquet format containing image bytes and rich detection metadata:

from datasets import load_dataset

dataset = load_dataset("VietPhong/kitti-yolo11n-robustness-benchmark", split="train", streaming=True)
sample = next(iter(dataset))

print("Original Image ID:", sample["original_image_id"])
print("Attack:", sample["attack_name"], "Severity:", sample["severity"])
print("PSNR (dB):", sample["psnr_db"], "SSIM:", sample["ssim"])
print("YOLO11n Clean Predictions:", sample["clean_predictions"])
print("YOLO11n Attacked Predictions:", sample["attacked_predictions"])

??? Reproduction & Provenance

Generated using the AdverTest Framework. All evaluations use strict FP32 baseline execution.