Dense Prediction and Depth Anything Shows 22% Accuracy in Chimpanzee Estimation

Researchers are tackling the laborious task of estimating chimpanzee populations using artificial intelligence. Tom Raynes and Otto Brookes, both from the University of Bristol, alongside Timm Haucke from the Massachusetts Institute of Technology, Lukas Bösch and Anne-Sophie Crunchant from the Wild Chimpanzee Foundation, and Hjalmar Kühl from the Senckenberg Museum of Natural History, Germany, present a novel approach integrating computer-based depth estimation into standard camera trap workflows. This study is significant because manually measuring animal-to-camera distances from extensive video footage is incredibly time-consuming, hindering large-scale ape conservation efforts. By combining two monocular depth estimation models with distance sampling strategies on a real-world dataset of 220 chimpanzee videos, the team demonstrate a viable alternative, achieving population estimates within 22% of traditional manual methods and paving the way for automated population monitoring.

Researchers tackled the labour-intensive process of manually measuring animal-to-camera distances, a crucial step in determining population density and abundance.

This study introduces and evaluates a computer vision pipeline integrating monocular depth estimation (MDE) directly into ecological workflows for great ape conservation. The team achieved this by combining two MDE models, Dense Prediction Transformers and Depth Anything, with various distance sampling strategies.
Using a dataset of 220 camera trap videos of wild chimpanzees, the models generated distance estimates, which were then used to infer population density and abundance. Comparative analysis against manually measured distances revealed that a calibrated DPT model consistently outperformed Depth Anything in both distance estimation accuracy and subsequent density and abundance inference.

However, the study unveils systematic biases in both models, particularly a tendency to overestimate detection distances and underestimate density and abundance in complex forest environments. Failures in animal detection across different ranges were identified as a primary limiting factor in estimation accuracy.

Despite these biases, the proposed MDE-driven camera trap distance sampling proved to be a viable and practical alternative to manual distance estimation. Experiments show the automated approach yields population estimates within 22% of those obtained using traditional manual methods. This work establishes a realistic pathway towards scalable and automated population modelling for great apes, offering a significant advancement in conservation efforts. The research opens possibilities for rapid and reliable data collection, enabling more timely and effective conservation action for these endangered primates.

Evaluating monocular depth estimation for chimpanzee distance sampling requires careful consideration

Scientists investigated automated methods for estimating great ape populations using camera trap footage. The study focused on replacing labour-intensive manual distance measurements with computer-based monocular depth estimation (MDE) pipelines. Researchers analysed 220 camera trap videos of a wild chimpanzee population to test this approach, combining two MDE models, Dense Prediction and Depth Anything, with various distance sampling strategies.

This integration aimed to generate accurate detection distance estimates, subsequently used to infer population density and abundance. The team calibrated and compared the performance of DPT and Depth Anything against manually derived ground-truth distances. Results demonstrated that DPT consistently outperformed Depth Anything in both distance estimation accuracy and downstream inference of density and abundance.

Despite this advantage, both models exhibited systematic biases, tending to overestimate detection distances and underestimate density and abundance in complex forest environments. Experiments revealed that failures in animal detection across different distance ranges significantly limited the overall estimation accuracy.

This work pioneered a practical MDE-driven camera trap distance sampling workflow, demonstrating its viability as an alternative to manual distance estimation. The proposed approach achieved population estimates within 22% of those obtained using traditional methods. Scientists employed MegaDetector and Segment Anything Model (SAM) to initially localise and segment chimpanzees within camera trap frames.

These segmented images were then processed using the MDE pipelines to generate depth maps, enabling the calculation of distances. The system delivers a streamlined pipeline, reducing analysis time from weeks of manual labour to hours of computation, and establishing a pathway towards scalable, automated population modelling for great apes.

Depth estimation from camera traps informs chimpanzee density and abundance in fragmented forests

Scientists investigated the potential of computer-based monocular depth estimation (MDE) to improve great ape population monitoring. The research team evaluated MDE pipelines integrated into camera trap workflows, focusing on a dataset of 220 chimpanzee videos. Combining two MDE models, Dense Prediction and Depth Anything, with various distance sampling strategies, they generated detection distance estimates to infer population density and abundance.

Experiments revealed that the calibrated DPT model consistently outperformed Depth Anything in both distance estimation accuracy and subsequent density and abundance inference. However, both models exhibited systematic biases, tending to overestimate detection distances and underestimate density and abundance compared to manual methods.

Data shows failures in animal detection across different distances significantly limited estimation accuracy. The proposed approach yielded population estimates within 22% of those obtained using traditional manual methods. Measurements confirm that MDE-driven camera trap distance sampling is a viable alternative to manual distance estimation.

Specifically, the team measured a consistent performance level, demonstrating a realistic pathway toward scalable and automated population modelling for great apes. Results demonstrate the framework’s ability to process frames containing calibration markers and habitat footage, localising and segmenting content of interest using MegaDetector and SAM.

These processed frames were then used to generate raw depth maps, ultimately yielding distance estimates for detected chimpanzees. The study highlights a practical pipeline, transforming raw camera trap video into population estimates, reducing labour-intensive manual processing from days or weeks to hours of computation.

Automated chimpanzee density estimation via monocular depth and camera traps offers a novel conservation tool

Scientists have demonstrated a viable pathway towards automating great ape population estimation using monocular depth estimation (MDE) integrated with camera trap distance sampling. The research involved analysing 220 chimpanzee camera trap videos, combining two MDE models, Dense Prediction and Depth Anything, with various distance sampling strategies to infer population density and abundance.

Comparative analysis revealed that the DPT model consistently outperformed Depth Anything in both distance estimation accuracy and subsequent density/abundance inference. However, both models exhibited systematic biases, tending to overestimate detection distances and underestimate population density relative to traditional manual methods.

Failures in animal detection across different distances were identified as a primary limitation to estimation accuracy. Despite these limitations, the automated approach yielded population estimates within 22% of those obtained using conventional manual techniques, significantly reducing time and labour requirements.

The study acknowledges that improved calibration, more robust animal localisation, and depth estimation models specifically tailored for complex forest environments are essential for operational deployment. Future research should focus on addressing these biases and detection failures to enhance the scalability and accuracy of automated bio-monitoring for great apes and other biodiversity assessments.