Official implementation of the paper in CVPR 2026.

Addressing the prohibitive costs of data annotation and the scarcity of sewer defect samples, we propose SFR-Net, a novel Multi-Label Zero-Shot Learning (ML-ZSL) framework. To mitigate "Alignment Ambiguity" in complex pipe environments, SFR-Net employs a three-stage paradigm: Representation Steering (RS) for scene adaptation, Multi-Granularity Evidence Fusion (MEF) for decoupled feature aggregation, and Generalized Relational Score Refining (GR) for transferring relational logic to unseen defects. Experiments on the Sewer-ML and WZ-Pipe datasets demonstrate that SFR-Net achieves state-of-the-art (SOTA) performance and significantly boosts zero-shot generalization.

a large-scale, multi-label benchmark dataset specifically designed for sewer pipe defect classification. It contains over 1.3 million images with 17 distinct defect categories. We select the five least-frequent categories as the unseen defects, ensuring no corresponding samples exist in the training set.

[Dataset] [Split]

a distinct dataset that employs different inspection standards from Sewer-ML. Specifically, the standards followed by WZ-Pipe feature more detailed and well-defined defect categories. It comprises approximately 60,000 samples across 17 categories, with the partition of unseen classes remaining consistent with that of Sewer-ML to ensure experimental comparability.

Download this dataset by submitting a form in the following link:

https://forms.gle/AJVLWeQDv1uGVnG87

Note on Dataset Versioning

The current release is a refined version of the dataset reported in the paper. We have performed a second round of re-annotation and removed low-quality or privacy-sensitive samples for better usability and safety.

Result on refined WZ-Pipe Dataset

Install the environment through conda:

conda env create -f environment.yml

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node=4 --master_addr=localhost --master_port=12355 \
main_mlzsl.py --config_file configs/sewerml.yml

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node=4 --master_addr=localhost --master_port=12355 \
main_mlzsl.py --config_file configs/sewerml.yml MODEL.LOAD True TEST.EVAL True TEST.WEIGHT sewerml_best.pth

python main_mlzsl.py --config_file configs/sewerml.yml MODEL.DIST_TRAIN False MODEL.LOAD True TEST.EVAL True TEST.WEIGHT sewerml_best.pth

If you find our work useful, please cite our paper:

@InProceedings{Chen_2026_CVPR,
    author    = {Chen, Zhao-Min and Huang, Xinjian and Ge, Yisu and Li, Yu},
    title     = {SFR-Net: Steering-Fusion-Refining Network in Multi-label Zero-Shot Sewer Defect Detection},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {41636-41645}
}

This repo benefits from RAM, CLIP and CLIP-Adapter. Thanks for their wonderful works.