DeepEthogram

DeepEthogram assigns predefined behavior labels to individual video frames. It's designed for neuroscience researchers measuring animal activity, and can also serve other projects that need action detection at frame level.

• Behavior timing: Researchers can measure scratching or courting, count licks, and identify reach onset times for comparison with neural activity.
• Model training: PyTorch training and prediction pipelines start from weights pretrained on Kinetics700 rather than requiring training from scratch.
• Desktop and cluster use: A graphical interface supports work on a local workstation, while a command line interface supports training on a compute cluster.
• Annotated datasets: The authors provide the datasets they collected and labeled for the project.

Local use requires a CUDA-capable NVIDIA GPU; AMD GPUs aren't supported. The project recommends at least 6 GB of VRAM, a CPU with four or more cores and an SSD. GPU performance and available VRAM affect processing speed and how much data fits into a training batch.

A Google Colab notebook provides cloud training for users without a suitable GPU. That route requires uploading data to Colab, rather than keeping training on the local workstation.

DeepEthogram is free for academic use under a custom license. Commercial use requires contacting Harvard's Office of Technology Development about licensing.