Research map / Label-efficient and robust learning / Self-supervised and contrastive

On-Device Self-Supervised Learning of Low-Latency Monocular Depth from Only Events

CVPR 2025 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Pretext-task visual pre-training · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Makes contrast-maximization self-supervised monocular depth from events fast enough to train on a drone by replacing batched PyTorch warping with per-event CUDA kernels, then shows online in-flight fine-tuning improves obstacle avoidance.

Problem

Self-supervised learning of event-based monocular depth via contrast maximization is computationally too expensive for real-time on-board learning on resource-constrained drones, and pre-training alone suffers from reality gaps.

Approach

Parallelize warping, bilinear splatting and their gradients over all accumulated events in custom CUDA kernels (one thread per event(s)) instead of padded batched PyTorch, removing padding and out-of-image warps, and combine this with a 430k-parameter ConvGRU depth/ego-motion network trained with a geometry consistency loss so it can fine-tune online during flight.

Key results

• On MVSEC (outdoor_day1, 10m cutoff), the proposed method reaches 2.25 m MAE versus 2.72 m for Zhu et al. 2019 among event-only self-supervised baselines; the dense variant reaches 1.96 m.
• The CUDA implementation gives roughly 100x runtime reduction and 2-5x lower memory versus batched PyTorch for common event rates.
• In flight, pre-trained plus online learning improves distance between pilot interventions by about 30% over pre-training only, and about 65% over training from scratch being near-zero obstacle avoidance.

Measured effects reported in the paper

Surprising findings

• Training from scratch produced no meaningful obstacle avoidance within the flight time, showing pre-training is essential for fast online adaptation.

Limitations

• Depth on DSEC falls short of supervised stereo baselines, and the metric scale must be estimated via a scaling factor.
• Only a quarter of the 640x480 camera resolution is used and the environment is artificially textured.
• Depth-based yaw control is attracted to corners, requiring pilot intervention; static-scene assumption underestimates moving objects.

Experimental setup

Datasets: UZH-FPV MVSEC DSEC

Benchmarks: MVSEC DSEC disparity benchmark real-world drone flight obstacle avoidance

Metrics: MAE RMSE RSAT distance between pilot interventions

Baselines: Zhu et al. (2019) unsupervised event depth Zhu et al. (2023) cross-modal consistency Cho et al. supervised stereo event depth DSEC baseline Training from scratch Ground-truth depth (RealSense)

Builds on: contrast maximization (Gallego et al.) Taming Contrast Maximization (Paredes-Vallés et al.) Zhou et al. unsupervised depth and ego-motion geometry consistency loss (Bian et al.) Lightweight event-based optical flow via iterative deblurring

Closest papers by meaning

• Self-Supervised Monocular Trained Depth Estimation Using Self-Attention and Discrete Disparity Volume (CVPR 2020 · similarity 0.85)
• Learning Depth from Past Selves: Self-Evolution Contrast for Robust Depth Estimation (AAAI 2026 · similarity 0.85)
• EventFly: Event Camera Perception from Ground to the Sky (CVPR 2025 · similarity 0.84)
• DERD-Net: Learning Depth from Event-based Ray Densities (NeurIPS 2025 · similarity 0.83)
• PPEA-Depth: Progressive Parameter-Efficient Adaptation for Self-Supervised Monocular Depth Estimation (AAAI 2024 · similarity 0.83)
• Synthetic-to-Real Self-supervised Robust Depth Estimation via Learning with Motion and Structure Priors (CVPR 2025 · similarity 0.83)