1
Bias-resistant distillationRebalance a 44M multi-source corpus, then distill DepthAnything v2-L.
Depth Anything Rethought for Tiny Models
DepthART
ACMMM 2026Scaling Foundation Monocular Depth to Tiny Models
🔥 New · September 2026
DepthART now supports MobileNetV4 backbones for earlier-generation, resource-constrained devices. MobileNetV4-M-slim-SPF pairs a slimmer encoder with Single-Path Pyramid Fusion for portable BPU/NPU deployment.
6.0M / 8.0MRelative-S / Metric-S parameters
0.92 msDepthART-S 224 · A6000 TensorRT FP16
1088.9DepthART-S 224 · A6000 FPS · TensorRT FP16
247.8 FPSOrin NX · TensorRT FP16
0.971 δ1DepthART-L · NYUD v2 zero-shot
Videos from ADVIO datasets, captured on iPhone
Relative Depth in the Wild
DepthART processes handheld portrait video while preserving fine boundaries and accurate depth.
RGB
DepthART-L · Relative
DepthART
Recent geometric foundation models have advanced monocular depth estimation, yet their benefits remain limited for tiny models. We present DepthART, a compact model designed for robust on-device depth estimation across diverse scenes. To address dataset-specific overfitting and unstable metric adaptation under camera shifts, DepthART combines bias-resistant data sampling with camera-conditioned fine-tuning that preserves the distilled encoder while adapting metric scale using camera intrinsics. These designs improve both cross-dataset generalization and metric depth prediction in capacity-constrained models.
Multi-platform Deployment
Accuracy vs Speed
RTX A6000PyTorch FP32Relative · NYUD v2
Additional comparison familiesAwaiting aligned A6000 FP32 model-only latency
DepthART encoderTinyViM-STinyViM-BTinyViM-LMobileNetV4-SMobileNetV4-MMobileNetV4-M-slim-SPF
Move across the plot to magnify nearby latency values.Comparison methods use affine-invariant δ1.Self-collected Image
Relative Depth
Scene 01
Left · Comparison method
Right · DepthART
YOLO26-L-DepthDepthART MobileNetV4-M-Slim-SPF · 448Web Video
Relative Depth
Busy city streetRobot deployment
Metric Point Cloud
DepthART-Metric-LFine-tuned on NYUD v2 only
Cafe · Scene 1Metric 3D reconstruction
Cite this work
DepthART
Accepted to ACM Multimedia 2026. The final citation and public model links will be added with the camera-ready release.
@inproceedings{depthart2026,
title = {DepthART: Scaling Foundation Monocular Depth to Tiny Models},
author = {Feng Xue and Wu Chen and Mingshuai Zhao and Guofeng Zhong and Anlong Ming and Haozhe Wang and Dianqiao Lei and Zhaowen Lin and Haiyang Zhang and Nicu Sebe},
booktitle = {ACM Multimedia},
year = {2026}
}