Leaderboard
Evaluation Metrics
- AUPR : pixel-wise Area Under Precision Recall curve
- FPR95 : pixel-wise False Positive Rate at a true positive rate of 95%
- sIoU gt : adjusted Intersection over Union averaged over all ground truth segmentation components
- PPV : predictive positive value (or precision) averaged over all predicted segmentation components
- mean F1 : component-wise F1-score averaged over different detection thresholds
Anomaly Track
Method
OoD Data
Pixel Level
AUPR
Component Level
sIoU gt
PPV
mean F1
DenseHybrid (DeepLabv3+)
42.05%
62.25%
36.90%
18.70%
11.32%
VLAD
92.94%
3.25%
71.58%
53.71%
65.40%
VLAD
82.48%
82.42%
66.99%
51.86%
61.14%
Con2MAV
90.00%
2.68%
59.10%
68.32%
69.38%
OodDINO+RPL
87.33%
7.83%
48.06%
52.41%
56.09%
OodDINO+RbA
85.64%
7.79%
46.21%
55.24%
54.90%
Obstacle Track
Method
OoD Data
Pixel Level
AUPR
Component Level
sIoU gt
PPV
mean F1
RAOS (training)
87.40%
0.28%
46.75%
46.60%
47.57%
RAOS (training free)
79.67%
0.81%
42.42%
32.58%
32.70%
DenseHybrid (DeepLabv3+)
80.79%
6.02%
48.48%
60.16%
55.59%
VLAD
78.70%
0.57%
36.42%
22.40%
24.36%
VLAD
76.43%
0.58%
42.47%
19.05%
23.63%
OodDINO+RbA
94.52%
0.05%
73.01%
80.44%
89.91%
OodDINO+RPL
94.05%
0.06%
67.72%
81.36%
86.53%
LostAndFound NoKnown
Method
OoD Data
Pixel Level
AUPR
Component Level
sIoU gt
PPV
mean F1
RAOS (training)
89.81%
0.95%
53.06%
57.86%
60.21%
RAOS (training free)
81.13%
3.23%
53.54%
46.20%
50.64%