CorrCLIPv2: Pre-generated Masks and Prebuilt Memory Banks

This repository holds the pre-generated region masks and the prebuilt memory banks used to reproduce CorrCLIPv2, the journal extension of CorrCLIP: Reconstructing Patch Correlations in CLIP for Open-Vocabulary Semantic Segmentation.

CorrCLIPv2 augments single-image open-vocabulary segmentation with cross-image semantic support retrieved from an offline memory bank built purely from unlabeled images — no annotation, no caption, no training. Downloading the artifacts here lets you skip both the mask generation and the bank construction step.

For the code, installation instructions and evaluation scripts, see the GitHub repository: https://github.com/zdk258/CorrCLIP (the CorrCLIPv2/ subdirectory).

Region masks

instance_mask.zip (110 M) — EntitySeg region masks for the ten standard benchmarks. Extract it into data/, which gives data/instance_mask/{voc,context,coco,ade,ade847,city}/ as instance_mask_path expects. With these in place, set mask_generator=None and no mask generator needs to be installed.

Prebuilt banks

Two banks are provided: coco, built from COCO train2017 and used for the main results, and monet, built from MONET(Z-Image) samples — fully synthetic and disjoint from every evaluation set.

One configuration needs exactly three files — a main store matching the CLIP scale, the DFN auxiliary store, and the FAISS index — so pick a row and download only those. CorrCLIPv2-B and -L share the auxiliary store and the index; the only difference between them is metaclip_b vs metaclip_l.

Configuration	Files to download	Total
CorrCLIPv2-L + coco (main results)	coco_metaclip_l.pt 4.2G · coco_dfnclip_b.pt 2.8G · coco_radiov3.faiss 190M	7.2G
CorrCLIPv2-B + coco	coco_metaclip_b.pt 2.8G · coco_dfnclip_b.pt 2.8G · coco_radiov3.faiss 190M	5.8G
CorrCLIPv2-L + monet (synthetic, disjoint)	monet_metaclip_l.pt 2.2G · monet_dfnclip_b.pt 1.4G · monet_radiov3.faiss 108M	3.7G
CorrCLIPv2-B + monet	monet_metaclip_b.pt 1.4G · monet_dfnclip_b.pt 1.4G · monet_radiov3.faiss 108M	3.0G

Put the .pt files in memory_bank/clip_embeddings/ and the .faiss file in memory_bank/vfm_embeddings/, keeping the filenames unchanged — the loader resolves each by name. Then set _clip_scale and _mb_bank in configs/base_config.py to match the row you chose and run as usual; vfm_type stays radiov3, the space these indexes were built in. For the no-DMCE ablation, set use_aux_model=False and skip the DFN store.

Each .pt store is an [N, D] fp16 tensor loaded fully into RAM at startup; its row index is the region id, and the FAISS index (IDMap2) returns the same ids, so search results index the stores directly.

Bank image lists

The image lists behind the disjoint banks are released for auditing — one source image path per line, the exact set each bank was built from: gpic (6.2 M), monet (2.8 M), sa1b (1.0 M). The coco bank needs no list: it is the full COCO train2017 image set.

Source data

The masks are derived from the public validation images of the corresponding benchmarks; the coco bank is derived from COCO train2017 and the monet bank from the MONET Z-Image synthetic subset. Only derived features are redistributed here, never the source images — the original datasets' own licenses and terms continue to apply to them. The disjoint-bank source images are available from SA-1B, GPIC, and MONET.

Citation

The CorrCLIPv2 journal extension is under review; please cite the conference paper for now:

@article{zhang2024corrclip,
  title={Corrclip: Reconstructing patch correlations in clip for open-vocabulary semantic segmentation},
  author={Zhang, Dengke and Liu, Fagui and Tang, Quan},
  journal={arXiv preprint arXiv:2411.10086},
  year={2024}
}