Studying animal behavior is crucial for understanding how different species and individuals interact with their surroundings. Video coding is preferred for collecting detailed behavioral data, but manually extracting information from extensive video footage is time-consuming. Likewise, manually coding animal behavior demands significant training for reliability.
Machine learning has emerged as a solution, automating data extraction and improving efficiency while maintaining reliability. It has successfully recognized species, individuals, and specific behaviors in videos, transforming behavioral research by tracking species in camera-trap footage and identifying animals in real time.
Yet, challenges remain in tracking nuanced behavior, especially in wild environments. While current tools excel in controlled settings, recent progress suggests the potential for expanding these techniques to diverse species and complex habitats. Combining machine learning methods, such as spatiotemporal action CNNs and pose estimation models, offers a holistic view of behavior over time.
In this context, a new paper was recently published in the Journal of Animal Ecology revolving around machine learning tools, particularly DeepLabCut, in analyzing behavioral data from wild animals, especially primates like chimpanzees and bonobos. It highlights the challenges faced in manually coding and extracting behavioral information from extensive video footage and the potential of machine learning to automate this process, thus significantly reducing time and improving reliability.
The paper details the use of DeepLabCut for analyzing animal behavior, citing various guides for installation and initial use, emphasizing the need for Python installation. It also discusses hardware requirements, including the recommendation for a GPU and the option to use Google Colaboratory. The GUI’s functionalities, limitations, and the need for loss graphs to gauge model training progress are covered. The extraction of video data from the Great Ape Dictionary Database and ethical considerations regarding data collection are highlighted.
Additionally, the paper outlines the video selection criteria, including visual ‘noise’ for diverse learning experiences, and the challenges in determining the required number of training frames based on data complexity. Model development, training sets, and video preparation methods are detailed, discussing limitations regarding frame marking time and hardware used. The performance assessment of the trained models, including comparisons between model-generated and human-labeled points, is explained, along with evaluations on test frames and novel videos.
The authors conducted experiments using DeepLabCut to develop and assess models for tracking the movements of wild chimpanzees and bonobos. They trained two models on different video frames, comparing their performance on both test frames (which contained some training data) and entirely new videos.
- Model 1 was trained on 1375 frames, while Model 2 used a larger set of 2200 frames, including input from a second human coder and data from an additional chimpanzee community.
- Key points on the primates in the video frames were marked to facilitate training.
- Both models were tested on frames used during training (test frames) and entirely new videos (novel videos) to assess their accuracy in tracking primate movements.
The evaluation of test frames revealed that both models exhibited enhanced accuracy in marking key points on video frames of wild chimpanzees compared to human coder variation. Model 2 consistently outperformed Model 1 across multiple body parts in these test frames. Additionally, when tested on novel videos, Model 2 displayed superior capabilities in detecting body points and accuracy across various body parts compared to Model 1. Despite these improvements, both models faced difficulties effectively linking detected points, resulting in tracking issues in specific videos.
The study revealed promising results in using DeepLabCut for tracking primate movements in natural settings. However, it highlighted the need for human intervention to correct tracking errors and the time-intensive nature of developing robust models through extensive training.
In conclusion, the paper demonstrates the potential of DeepLabCut and machine learning in automating the analysis of wild primate behavior. While it marks significant progress in tracking animal movements, challenges persist, notably the need for human intervention in error correction and the time-intensive model development process. These findings highlight the transformative impact of machine learning in behavioral research while underscoring the ongoing need for refinement in tracking systems for nuanced behavior in natural settings.
Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter. Join our 35k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.
Mahmoud is a PhD researcher in machine learning. He also holds a
bachelor’s degree in physical science and a master’s degree in
telecommunications and networking systems. His current areas of
research concern computer vision, stock market prediction and deep
learning. He produced several scientific articles about person re-
identification and the study of the robustness and stability of deep