From relative to metric: Calibrating AI-based monocular depth learning models for distance sampling in wildlife monitoring applications

Distance sampling is a widely adopted statistical method for estimating animal population density by measuring distances between predefined sampling points or transects and detected individuals in a study area. These measurements enable modeling of detection probability, facilitating accurate population assessments. Camera Trap Distance Sampling (CTDS) is an innovative adaptation that replaces human observers with a network of stationary camera traps to record animal presence. This automation expands the method’s potential for quantitative studies but introduces challenges in processing large volumes of video data. A key requirement of CTDS is the ability to estimate the distance between animals and the camera, thereby shifting the methodological challenge from field-based human measurement to computational depth estimation. Recent advancements in computer vision have significantly improved monocular depth estimation (MDE); existing approaches often rely on extensive manual annotations and complex, multi-stage calibration or training processes for estimating metric distances. Beyond camera-trap distance sampling, metric monocular distance estimates enable analyses such as body-size/height estimation, 3D scene reconstruction, and quantification of behavior and movement (e.g., movement rates and encounter models). In CTDS, the application of MDE shifts distance measurement from manual fieldwork to automated per-frame inference; we use CTDS as a primary motivating application. We propose a lightweight, model-agnostic calibration method that converts off-the-shelf relative MDE predictions into metric distances using as few as three reference points with known distances. Our per-image/per-frame least-squares calibration enables metric distance estimation from a single image and is efficient enough for near real-time use on resource-constrained hardware. We evaluate seven state-of-the-art MDE models and show that image-based models provide the best practical trade-off for per-frame processing. In contrast, diffusion-based and video depth methods are prohibitively slow for this setting. Our pipeline, encompassing animal detection, depth extraction, calibration, and distance estimation, achieves an average processing time of four seconds. Among all evaluated depth-estimation models, Depth Anything V2 consistently delivered the highest accuracy. Our point-based calibration method for animal distance sampling is the top performer, achieving a median absolute relative error of 0.125 for animal distance estimates, surpassing AUtomated DIstance esTimation (AUDIT), the current state-of-the-art model, which has a median absolute relative error of 0.433.