Depth Anything V2 Small — ONNX

ONNX export of depth-anything/Depth-Anything-V2-Small-hf, a monocular depth estimation model. Predicts relative inverse depth from a single RGB image — larger values are closer to the camera.

Mirrored for use with inference4j, an inference-only AI library for Java.

Original Source

Repository: Depth Anything (ONNX by onnx-community)
License: apache-2.0

Usage with inference4j

try (DepthAnythingEstimator estimator = DepthAnythingEstimator.builder()
        .modelId("inference4j/depth-anything-v2-small")
        .build()) {
    DepthMap depth = estimator.estimate(Path.of("photo.jpg"));
    System.out.println("Depth at centre: " + depth.at(depth.width() / 2, depth.height() / 2));
}

Model Details

Property	Value
Architecture	DINOv2 ViT-Small backbone + DPT decoder
Task	Monocular depth estimation
Parameters	~25M
Input	`[1, 3, height, width]` — NCHW float32, RGB
Normalization	ImageNet mean `[0.485, 0.456, 0.406]`, std `[0.229, 0.224, 0.225]`
Input sizing	518 on the longer edge, aspect preserved, both dimensions rounded up to a multiple of 14 (ViT patch size), bicubic resampling
Output name	`predicted_depth`
Output	`[1, height, width]` — rank 3, no channel dimension
Output semantics	Relative inverse depth; larger is closer. Scale is per-image, not metric.
Original framework	PyTorch (HuggingFace Transformers)

License

This model is licensed under the Apache 2.0 License. Original model by Depth Anything, ONNX export by onnx-community.

Preprocessing

Resize so the longer edge is 518, preserving aspect ratio (bicubic)
Round both dimensions up to a multiple of 14 (the ViT patch size)
ImageNet normalization: (pixel / 255 - mean) / std
- mean = [0.485, 0.456, 0.406]
- std = [0.229, 0.224, 0.225]
NCHW layout: [1, 3, H, W]

Postprocessing

The output is relative inverse depth, so values are only comparable within a single image. To render a depth map, normalize to [0, 1] against the per-image min and max, then resize back to the original image dimensions.

Original Paper

Yang, L., Kang, B., Huang, Z., Zhao, Z., Xu, X., Feng, J., & Zhao, H. (2024). Depth Anything V2. NeurIPS 2024. arXiv:2406.09414

inference4j/depth-anything-v2-small