qbit-glitch/cityscapes-pseudo-labels

Cityscapes Unsupervised Panoptic Pseudo-Labels

Pseudo-labels for unsupervised panoptic segmentation on Cityscapes, generated using overclustered k-means semantics + depth-guided instance splitting.

Contents

Pseudo-Labels

Directory	Description	Files	Format
pseudo_semantic_raw_k80/	Overclustered k=80 semantic labels	~3.5K PNGs + centroids.npz	PNG (values 0-79), train/val split
cups_pseudo_labels_depthpro_tau020/	CUPS-format combined labels (DepthPro tau=0.20)	~9K files	semantic.png + instance.png + .pt per image

Pre-trained Weights

File	Description	Size
weights/dinov3_vitb16_official.pth	DINOv3 ViT-B/16 backbone (Meta official format)	327MB
weights/cups.ckpt	CUPS Cascade Mask R-CNN checkpoint	916MB

Download

# Everything
huggingface-cli download qbit-glitch/cityscapes-pseudo-labels --repo-type dataset --local-dir ./data/

# Just pseudo-labels (no weights)
huggingface-cli download qbit-glitch/cityscapes-pseudo-labels --repo-type dataset --local-dir ./data/ \
  --include "pseudo_semantic_raw_k80/*" "cups_pseudo_labels_depthpro_tau020/*"

# Just weights
huggingface-cli download qbit-glitch/cityscapes-pseudo-labels --repo-type dataset --local-dir ./data/ \
  --include "weights/*"

# Just DINOv3 backbone
huggingface-cli download qbit-glitch/cityscapes-pseudo-labels --repo-type dataset --local-dir ./data/ \
  --include "weights/dinov3_vitb16_official.pth"

Pipeline

1. Semantic: DINOv2 ViT-B/14 features -> k-means (k=80) overclustering -> 80-class pseudo-semantics
2. Instance: DepthPro monocular depth -> Sobel gradients -> threshold (tau=0.20, A_min=1000) -> instance masks
3. CUPS format: semantic.png (mapped to 27 CAUSE classes) + instance.png + .pt (metadata) per image

Metrics (Cityscapes val, 27-class CAUSE + Hungarian matching)

Component	PQ	PQ_stuff	PQ_things
k=80 semantics + DepthPro instances	28.40	32.08	23.35
CUPS Stage-2 (trained on these labels)	30.26	31.29	28.50

Citation

If you use these pseudo-labels, please cite CUPS (CVPR 2025).