P2A: Transforming Proposals to Anomaly Masks (ICPR 2024)

Authors: Huachao Zhu, Zhichao Sun, Zelong Liu, Yongchao Xu,

Figure 1: Qualitative comparison on ACDC. The yellow box represents the densely partitioned VP region. Our model produces more accurate results for both distant tiny hard samples near the VP and occluded fast-moving close targets.

Installation

Please refer to get_started.md for installation.

Here is an example:

conda create -n p2a python=3.9
conda activate p2a
conda install pytorch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 pytorch-cuda=11.8 -c pytorch -c nvidia
pip install -U openmim
mim install mmengine
mim install "mmcv>=2.0.0"
mim install mmdet
pip install -v -e .

Datasets

• Cityscapes: consists of inlier classes.
• MS COCO: helps fine-tune the model on ood-objects.
• Road Anomaly Datasets: RoadAnomaly can be downloaded at link. FS Static validation can be downloaded according to Mask2Anomaly. SMIYC RoadAnomaly21 can be downloaded at link.

Quantitative Results

The results for various methods on Road Anomaly, FS L&F and FS Static are shown as below.

Methods	Road Anomaly AUC ↑	Road Anomaly AP ↑	Road Anomaly FPR95 ↓	FS L&F AUC ↑	FS L&F AP ↑	FS L&F FPR95 ↓	FS Static AUC ↑	FS Static AP ↑	FS Static FPR95 ↓
SynthCP	88.34	6.54	45.95	89.90	23.22	34.02	76.08	24.86	64.69
SML	81.96	25.82	49.74	96.88	36.55	14.53	96.69	48.67	16.75
Meta-OoD	-	-	-	93.06	41.31	37.69	97.56	72.91	13.57
SynBoost-WR38	81.91	38.21	64.75	96.21	60.58	31.02	95.87	66.44	25.59
MOoSe	-	43.59	32.12	-	-	-	-	-	-
PEBAL	87.63	45.10	44.58	98.96	58.81	4.76	99.61	92.08	1.52
ATTA	92.11	59.05	33.59	99.05	65.58	4.48	99.66	93.61	1.15
Mask2Anomaly	96.57	79.70	13.45	95.41	69.46	9.31	98.35	90.54	1.98
RbA	97.99	85.42	6.92	98.62	70.81	6.30	98.96	75.43	3.52
cDNP	-	85.6	9.8	-	-	-	-	-	-
P2A (ours)	98.40	89.42	5.95	97.24	65.15	13.98	99.66	96.93	0.11

The results for various methods on SMIFC RoadAnomaly21 are shown as below.

Methods	AP ↑	FPR95 ↓	SIoU gt ↑	PPV ↑	mean F1 ↑
Image Resynthesis	52.28	25.93	39.68	10.95	12.51
SML	46.8	39.5	26.0	24.7	12.2
SynBoost	56.44	61.86	34.68	17.81	9.99
Void Classifier	36.61	63.49	21.14	22.13	6.49
DenseHybrid	77.96	9.81	54.17	24.13	31.08
PEBAL	49.14	40.82	38.88	27.20	14.48
Mask2Anomaly	88.72	14.63	55.28	51.68	47.16
RbA	90.9	11.6	55.7	52.1	46.8
cDNP	88.90	11.42	50.44	29.04	28.12
ATTA	67.04	31.57	44.58	29.55	20.64
NFlowJS	56.92	34.71	36.94	18.01	14.89
P2A (ours)	91.5	8.9	55.5	52.2	53.4

Acknowledgement

We use some codes from repositories including ControlNet, Meta-OOD and Mask2Anomaly. We build our codes on mmsegmentation.