The Multi-Agent Behavior Dataset: Mouse Dyadic Social Interactions

Abstract

Multi-agent behavior modeling aims to understand the interactions that occur between agents. We present a multi-agent dataset from behavioral neuroscience, the Caltech Mouse Social Interactions (CalMS21) Dataset. Our dataset consists of trajectory data of social interactions, recorded from videos of freely behaving mice in a standard resident-intruder assay. To help accelerate behavioral studies, the CalMS21 dataset provides benchmarks to evaluate the performance of automated behavior classification methods in three settings: (1) for training on large behavioral datasets all annotated by a single annotator, (2) for style transfer to learn inter-annotator differences in behavior definitions, and (3) for learning of new behaviors of interest given limited training data. The dataset consists of 6 million frames of unlabeled tracked poses of interacting mice, as well as over 1 million frames with tracked poses and corresponding frame-level behavior annotations. The challenge of our dataset is to be able to classify behaviors accurately using both labeled and unlabeled tracking data, as well as being able to generalize to new settings.

Listing 1:

Json file format.

Figure 8:. Summary of Tasks.

Visual summary of datasets, tasks, and evaluations for the three tasks defined in CalMS21.

Figure 9:. Challenge Top-1 Model on Tasks 1 & 2.

Figure 10:. Challenge Top-1 Model on Task 3.

This model follows the MS-G3D Architecture [37]. “TCN”, “GCN”, “MS-” respectively denotes temporal and graph convolution blocks (STGC), and multi-scale aggregation.

Figure 11:

Learned Annotator Matrix from Top-1 Model on Tasks 1 & 2. The matrix is initialized as a diagonal matrix and each embedding dimension corresponds to the correponding annotator id. The lines on the top of the matrix represent the results from hierarchical clustering.

Figure 1:. Overview of behavior classification.

A typical behavior study starts with extraction of tracking data from videos. We show 7 keypoints for each mouse, and draw the trajectory of the nose keypoint. The goal of the model is to classify each frame (30Hz) to one of the behaviors of interest from domain experts.

Figure 2:. Behavior classes and annotator variability.

A. Example frames showing some behaviors of interest. B. Domain expert variability in behavior annotation, reproduced with permission from [55]. Each row shows annotations from a different domain expert annotating the same video data.

Figure 3:. Pose keypoint definitions.

Illustration of the seven anatomically defined keypoints tracked on the body of each animal. Pose estimation is performed using MARS [55].

Figure 4:. Available data for each task in our challenge.

Our dataset consists of a large set of unlabeled videos alongside a set of annotated videos from one annotator. Annot 1, 2, 3, 4, 5 are different domain experts, whose annotations for attack, mount, and investigation are used in Task 2. Bottom row shows new behaviors used in Task 3.

Figure 5:. Sequence Classification Setup.

Sequence information from past, present, and future frames may be used to predict the observed behavior label on the current frame. Here, we show a 1D convolutional neural network, but in general any model may be used.

Figure 6:. Baseline models.

Different baseline setups we evaluated for behavior classification. The input frame coloring follows the same convention as Figure 5: past frames in green, current frame in orange, and future frames in cyan.

Figure 7:

Example of errors from a sequence of behaviors from Task 1.