YOLO26 - Depth Estimation

YOLO26 is the latest addition to the Ultralytics YOLO model series. The depth estimation variant predicts a metric depth value in meters for every pixel, producing a dense monocular depth map from a single RGB image.

Overview

YOLO26 for depth estimation pairs the efficient YOLO26 backbone with a dense log-depth prediction head. Key features include:

• Metric depth in meters - The log-depth head predicts unbounded absolute depth (~0.02-150 m), unlike relative-depth models such as Depth Anything.
• Efficient YOLO26 backbone - Shares the NMS-free, DFL-free YOLO26 architecture for fast inference and broad edge compatibility.
• Broadly pre-trained checkpoints - Public weights trained on ~2.19M indoor and outdoor images, evaluated on NYU Depth V2, available across all model sizes.
• 768×768 inference - Checkpoints are trained and exported at 768×768; the head predicts at input/4 resolution and upsamples inside the exported graph.
• Multiple model sizes - From nano to extra-large variants.

Visit yolodepth.com for an overview of how to use YOLO26 depth estimation in your Roboflow project.

License

AGPL-3.0

Commercial Licensing

• AGPL-3.0: Free for open-source projects. Requires derivative works to be open-sourced.
• Paid Roboflow customers: Automatically get access to use any YOLO26 models trained on or uploaded to the Roboflow platform for commercial use.
• Free Roboflow customers: Can use YOLO26 via the serverless hosted API, or commercially self-hosted with a paid plan.

Learn more: Roboflow Licensing | YOLO26 License Details

Pre-trained Model IDs

Public YOLO26 depth estimation checkpoints are trained at 768×768 and are open access (no API key required).

Model Size	768×768
Nano	yolo26n-depth-768
Small	yolo26s-depth-768
Medium	yolo26m-depth-768
Large	yolo26l-depth-768
Extra-Large	yolo26x-depth-768

Supported Backends

Backend	Extras Required
onnx	onnx-cpu, onnx-cu12, onnx-cu118, onnx-jp6-cu126
torch-script	torch-cpu, torch-cu118, torch-cu124, torch-cu126, torch-cu128, torch-jp6-cu126
trt	trt10

Roboflow Platform Compatibility

Feature	Supported
Training	❌ Not yet available
Upload Weights	❌ Not yet available
Serverless API (v2)	✅ Deploy via hosted API
Edge Deployment (Jetson)	✅ Deploy on NVIDIA Jetson devices
Self-Hosting	✅ Deploy with inference-models

Usage Example

import cv2
import numpy as np
from inference_models import AutoModel

# Load a public checkpoint (open access, no API key required)
model = AutoModel.from_pretrained("yolo26n-depth-768")
image = cv2.imread("path/to/image.jpg")

# Run inference - depth in meters at the original image resolution
results = model(image)
depth_meters = results[0]  # torch.Tensor, (H x W), float32

print(f"Depth range: {depth_meters.min():.2f}m - {depth_meters.max():.2f}m")

# Colormap for visualization (near = bright, matching Depth Anything renders)
depth = depth_meters.cpu().numpy()
normalized = (depth.max() - depth) / (depth.max() - depth.min())
colored = cv2.applyColorMap((normalized * 255).astype(np.uint8), cv2.COLORMAP_VIRIDIS)
cv2.imwrite("depth_result.jpg", colored)

Output Format

The model returns a list of torch.Tensor depth maps, one per input image:

Field	Type	Description
results[i]	torch.Tensor	Metric depth in meters for each pixel (H x W, float32), at the original image resolution

Metric vs relative depth

AutoModel returns absolute metric depth in meters. When served through the Roboflow hosted API or the depth_estimation@v1 workflow block, the output is converted to the per-image ordinal-depth contract also used for Depth Anything (1.0 = nearest, 0.0 = farthest). The models therefore share the same output shape, range, and near-to-far ordering, but intermediate values are not geometrically equivalent or directly comparable across model families without calibration. The metric scale is not exposed on that path.

Choosing between YOLO26 depth and Depth Anything

YOLO26 predicts at input/4 native resolution and is substantially faster at a matched budget, with metric calibration; Depth Anything V3 decodes at full network resolution, producing sharper relative-depth maps at higher compute cost. Prefer YOLO26 for speed and, when using AutoModel directly, absolute distances. Prefer Depth Anything for edge fidelity in relative maps. The hosted API and workflow block normalize both families and do not expose YOLO26's absolute distances.