PatchRefinerV2

Fast and Lightweight Real-Domain High-Resolution
Metric Depth Estimation

Zhenyu Li, Wenqing Cui, Shariq Farooq Bhat, Peter Wonka.
KAUST

✨ NEWS

• 2026-02-17: PRV2 is accepted to ICLR 2026. Finish releasing everything.
• 2025-01-05: Release codes. Pretrained models are coming soon.

---

Installation

conda env create -n patchrefinerv2 --file environment.yml
conda activate patchrefinerv2

---

Data Preparation

UnrealStereo4K (U4K)

Download the dataset and place (or symlink) it under data/u4k/. The expected directory structure is:

data/u4k/
├── splits/
│   ├── val.txt
│   ├── test_in.txt
│   └── test_out.txt
└── ... (images and depth maps)

The split files are provided in splits/u4k/. You can symlink them:

ln -s $(pwd)/splits/u4k $(pwd)/data/u4k/splits

Cityscapes

Download the Cityscapes dataset and place (or symlink) it under data/cityscapes/. The split file data/cityscapes/val.txt listing relative image paths is required, which is also provided in splits/cityscapes/

---

Checkpoints

Download the pretrained checkpoints from here:

Model	Dataset	Config
PatchRefinerV2-U4K	UnrealStereo4K	configs/patchrefinerv2_u4k.yaml
PatchRefinerV2-CS	Cityscapes	configs/patchrefinerv2_cityscapes.yaml

Place checkpoints anywhere and pass the path via --ckp-path.

---

Inference

All inference is launched via tools/dist_test.sh, which automatically sets PYTHONPATH and spawns the distributed launcher.

bash tools/dist_test.sh <CONFIG> <NUM_GPUS> --ckp-path <CHECKPOINT> [OPTIONS]

Benchmark Evaluation

UnrealStereo4K

bash tools/dist_test.sh configs/patchrefinerv2_u4k.yaml 1 \
    --ckp-path /path/to/u4k_checkpoint.pth \
    --save \
    --work-dir ./output/u4k

Cityscapes

bash tools/dist_test.sh configs/patchrefinerv2_cityscapes.yaml 1 \
    --ckp-path /path/to/cs_checkpoint.pth \
    --image-raw-shape 1024 2048 \
    --patch-split-num 4 4 \
    --cai-mode m1 \
    --save \
    --work-dir ./output/cityscapes

Custom Image Inference

Provide a directory of images via --cfg-options and use --test-type general:

bash tools/dist_test.sh configs/patchrefinerv2_u4k.yaml 1 \
    --ckp-path /path/to/u4k_checkpoint.pth \
    --test-type general \
    --cfg-options general_dataloader.dataset.rgb_image_dir='/path/to/images' \
    --cai-mode r64 \
    --image-raw-shape 2160 3840 \
    --patch-split-num 4 4 \
    --save \
    --work-dir ./output/custom

---

Key Arguments

Argument	Description	Default
--ckp-path	Path to the .pth checkpoint file	''
--test-type	normal for benchmark datasets, general for custom images	normal
--cai-mode	Tiling mode: m1 (non-overlapping grid), m2 (shifted tiles), rN (N random patches, e.g. r64)	m1
--image-raw-shape H W	Full resolution of the input images (height width)	2160 3840
--patch-split-num R C	Number of patch rows and columns	4 4
--process-num	Batch size for patch inference (reduce if OOM)	4
--save	Save predicted depth maps and colorized visualizations	off
--work-dir	Output directory for logs and predictions	derived from --ckp-path
--cfg-options	Override any config value, e.g. key=value	—
--gray-scale	Use grayscale colormap for visualizations	off

Tiling Modes

• m1: Non-overlapping grid. Fast and memory-efficient.
• m2: Shifted-grid (overlapping) tiles. Better boundary handling.
• rN: N random patches sampled from the m2 set (e.g. r64). Good balance of coverage and speed.

Patch Split Constraints

The image dimensions must satisfy:

• height divisible by 2 × patch_split_num[0]
• width divisible by 2 × patch_split_num[1]

For example, with --image-raw-shape 2160 3840 --patch-split-num 4 4:

• 2160 / (2×4) = 270 ✓
• 3840 / (2×4) = 480 ✓
• Check here for more details.

---

Multi-GPU Inference

Pass the number of GPUs as the second argument to dist_test.sh:

bash tools/dist_test.sh configs/patchrefinerv2_u4k.yaml 4 \
    --ckp-path /path/to/checkpoint.pth \
    --save \
    --work-dir ./output/u4k

---

Citation

If you find our work useful for your research, please consider citing the paper

@article{li2025patchrefinerv2,
    title={PatchRefiner V2: Fast and Lightweight Real-Domain High-Resolution Metric Depth Estimation}, 
    author={Li, Zhenyu and Cui, Wenqing and Bhat, Shariq Farooq and Wonka, Peter},
    journal={arXiv preprint arXiv:2501.01121},
    year={2025}
}