README.md

---
tags:
  - open-vocabulary
  - dense-prediction
  - clip
  - cvpr-2025
  - tpami
pretty_name: DeCLIP TPAMI Release
---

# DeCLIP: Decoupled Learning for Open-Vocabulary Dense Perception

This is the **primary asset mirror** for the TPAMI release of **DeCLIP** (CVPR 2025 accepted). It hosts training checkpoints, pretrained backbones, a full source-code mirror, and analysis scripts.

- 📄 **Paper**: DeCLIP (CVPR 2025)
- 💻 **Code (canonical)**: [github.com/xiaomoguhz/DeCLIP_private](https://github.com/xiaomoguhz/DeCLIP_private) — training/evaluation scripts and setup instructions live there
- 🤗 **Sibling HF repos**: see [Repository Family](#repository-family) below

## What's in this repo

The directory layout on HuggingFace mirrors the [code repository](https://github.com/xiaomoguhz/DeCLIP_private) one-to-one:

```
DeCLIP-TPAMI/
├── src/                # source code mirror: training/, open_clip/, diffusion_model/, segment_anything/
├── scripts/            # training and evaluation scripts
├── tools/              # text embeddings, KNN, segmentation utilities
├── metadata/           # precomputed text embeddings (.npy)
├── third_party/        # ODISE, Mask2Former, TinyCLIP
├── downstream/         # CLIPSelf (F-ViT), ProxyCLIP_TPAMI, SegEarth-OV, mmsegmentation
├── analysis/           # visualization, failure cases, robustness, decoupling analysis, ...
├── deployment/         # TensorRT inference and quantization
├── experimental/       # CSAUpsampler, SimFeatUp, environment setup
├── demo_images/
│
├── checkpoints/        # Pretrained EVA-CLIP backbones
│   ├── EVA02_CLIP_B_psz16_s8B.pt       (300 MB)
│   └── EVA02_CLIP_L_336_psz14_s6B.pt   (856 MB)
│
└── logs/               # All training runs with trained DeCLIP / DeCLIP+ / Integrated / Ablation checkpoints
    ├── DeCLIP_EVA-B_DINOv2-B_560/checkpoints/
    ├── declip_plus_seg/epoch_6.pt
    ├── Integrated_EVA-B_DINOv2-B_560_2loss/checkpoints/
    ├── Integrated_EVA-B_DINOv2-B_560_3loss/checkpoints/
    ├── Integrated_EVA-L_DINOv2-L_336/checkpoints/
    ├── Ablation_JEPA-GSC_EVA-B_DINOv2-B_csa_560/checkpoints/
    ├── Ablation_JEPA-GSC_EVA-L_DINOv2-B_csa_336/checkpoints/
    ├── Ablation_SAM-GSC_EVA-B_DINOv2-B_csa_560/checkpoints/
    └── Ablation_SAM-GSC_EVA-L_DINOv2-B_csa_560/checkpoints/
```

## Usage

```bash
# Download the entire tree (~65 GB)
huggingface-cli download xiaomoguhzz/DeCLIP-TPAMI \
    --repo-type dataset --local-dir ./DeCLIP_TPAMI

# Download only one training run
huggingface-cli download xiaomoguhzz/DeCLIP-TPAMI --repo-type dataset \
    --include "logs/DeCLIP_EVA-B_DINOv2-B_560/checkpoints/*" \
    --local-dir ./DeCLIP_TPAMI

# Download only the pretrained EVA-CLIP backbones
huggingface-cli download xiaomoguhzz/DeCLIP-TPAMI --repo-type dataset \
    --include "checkpoints/EVA02*" \
    --local-dir ./DeCLIP_TPAMI
```

For full training and evaluation instructions, please refer to the [code repository README](https://github.com/xiaomoguhz/DeCLIP_private).

## Repository Family

This repo is the canonical release, but three earlier sibling repos preserve separable concerns from the project timeline. Together they form the complete TPAMI release:

| HF Repo | Type | Role | Created |
|---------|------|------|---------|
| **`xiaomoguhzz/DeCLIP-TPAMI`** (this repo) | dataset | Source code mirror, training logs, analysis scripts | 2026-04 |
| [`xiaomoguhzz/DeCLIP-TPAMI-data-2026-01`](https://huggingface.co/datasets/xiaomoguhzz/DeCLIP-TPAMI-data-2026-01) | dataset | COCO / LVIS / RefCOCO annotation bundles | 2026-01 |
| [`xiaomoguhzz/DeCLIP-TPAMI-weights-2026-01`](https://huggingface.co/xiaomoguhzz/DeCLIP-TPAMI-weights-2026-01) | model | F-ViT OV-COCO/LVIS detectors + SAM / TinyCLIP / CLIPSelf baselines | 2026-01 |
| [`xiaomoguhzz/DeCLIP-TPAMI-extras-2026-02`](https://huggingface.co/datasets/xiaomoguhzz/DeCLIP-TPAMI-extras-2026-02) | dataset | Archived analysis outputs and visualization packs | 2026-02 |

Ten legacy per-run model repos from mid-2025 are also retained for provenance — see the [code repository README](https://github.com/xiaomoguhz/DeCLIP_private#legacy-per-run-model-repos) for the full list. They do not correspond to the numbers reported in the paper; always use `DeCLIP-TPAMI` for the canonical release.

## Acknowledgement

DeCLIP builds upon [CLIPSelf](https://github.com/wusize/CLIPSelf), [ClearCLIP](https://github.com/mc-lan/ClearCLIP), [CAT-Seg](https://github.com/cvlab-kaist/CAT-Seg), [EVA-CLIP](https://github.com/baaivision/EVA/tree/master/EVA-CLIP), and [OpenCLIP](https://github.com/mlfoundations/open_clip).

## Citation

```bibtex
@article{declip2025,
  title   = {DeCLIP: Decoupled Learning for Open-Vocabulary Dense Perception},
  author  = {...},
  journal = {IEEE TPAMI (accepted)},
  year    = {2025}
}
```