The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
VIPER-N — High-Resolution (1920×1080) Driving Instance Segmentation Benchmark with Noisy Labels
VIPER-N provides calibrated noisy COCO-format instance segmentation annotations (Easy, Medium, Hard) for high-resolution (1920×1080) ego-centric urban driving scenes from the VIPER dataset, as introduced in:
Sfumato Supervision: Learning Instance Segmentation with Noisy Labels
Moshe Kimhi$^{1}$, Ehud Rivlin$^{1}$, Chaim Baskin$^{2}$
$^{1}$Technion – Israel Institute of Technology · $^{2}$Ben-Gurion University of the Negev
Project Page (mkimhi.github.io/Sfumato) | Paper (arXiv:2406.10891) | Code (github.com/mkimhi/noisy_labels)
Visual Overview of VIPER-N (1920×1080)
Comparison between Clean Ground-Truth driving masks and our calibrated procedural noise tiers (VIPER-N Medium and VIPER-N Hard), illustrating positional drift, polygon simplification, and super-category confusion on vehicles, pedestrians, and thin street infrastructure.
What’s Included in Kimhi/viper-n
| File Path | Tier / Split | Mean Mask IoU | Mean Boundary IoU (BIoU) |
Corrupted Instance BIoU |
Semantic / Omission Rate |
|---|---|---|---|---|---|
viper_n/train_clean.json |
Clean Reference Training Split (1920×1080) |
1.000 |
1.000 |
1.000 |
0% / 0% |
viper_n/val_clean.json |
Fixed Clean Reference Validation Split (1920×1080) |
1.000 |
1.000 |
1.000 |
0% / 0% |
viper_n/easy_seed1.json |
VIPER-N Easy (calibrated to human noise) | 0.712 |
0.708 |
0.583 |
5.0% super-cat / 5.0% omit |
viper_n/medium_seed1.json |
VIPER-N Medium (calibrated to weak prompts) | 0.528 |
0.522 |
0.444 |
8.0% super-cat / 8.0% omit |
viper_n/hard_seed1.json |
VIPER-N Hard (severe contour & semantic stress) | 0.379 |
0.369 |
0.334 |
12.0% super-cat / 12.0% omit |
benchmark/annotations/instances_{train,val}2017.json |
Legacy v1 VIPER-N annotations (preserved for compatibility) | -- | -- | -- | -- |
Quickstart: Pairing Kimhi/viper-n with Kimhi/viper Images
Download the images from Kimhi/viper and the noisy annotations from Kimhi/viper-n:
from huggingface_hub import snapshot_download
from pycocotools.coco import COCO
# 1. Download VIPER images + clean annotations
viper_root = snapshot_download("Kimhi/viper", repo_type="dataset")
images_train_dir = f"{viper_root}/images/train"
images_val_dir = f"{viper_root}/images/val"
# 2. Download VIPER-N noisy benchmark tiers
viper_n_root = snapshot_download("Kimhi/viper-n", repo_type="dataset")
# 3. Load any VIPER-N tier (Easy / Medium / Hard) and evaluate on val_clean.json
coco_train_med = COCO(f"{viper_n_root}/viper_n/medium_seed1.json")
coco_val_clean = COCO(f"{viper_n_root}/viper_n/val_clean.json")
print("VIPER-N Medium train annotations:", len(coco_train_med.anns))
print("VIPER Clean val annotations:", len(coco_val_clean.anns))
Citation
@article{kimhi2026sfumato,
title = {Sfumato Supervision: Learning Instance Segmentation with Noisy Labels},
author = {Kimhi, Moshe and Rivlin, Ehud and Baskin, Chaim},
journal = {arXiv preprint arXiv:2406.10891},
url = {https://arxiv.org/abs/2406.10891},
year = {2026}
}
- Downloads last month
- 63