MSGL-Transformer: A Multi-Scale Global-Local Transformer for Rodent Social Behavior Recognition

Muhammad Imran Sharif; Doina Caragea

cs.CVApr 8, 2026

MSGL-Transformer: A Multi-Scale Global-Local Transformer for Rodent Social Behavior Recognition

Authors: Muhammad Imran Sharif, Doina Caragea

Organizations: Department of Computer Science, Kansas State University, Manhattan, KS, 66506, USA.

Abstract

Recognition of rodent behavior is important for understanding neural and behavioral mechanisms. Traditional manual scoring is time-consuming and prone to human error. We propose MSGL-Transformer, a Multi-Scale Global-Local Transformer for recognizing rodent social behaviors from pose-based temporal sequences. The model employs a lightweight transformer encoder with multi-scale attention to capture motion dynamics across different temporal scales. The architecture integrates parallel short-range, medium-range, and global attention branches to explicitly capture behavior dynamics at multiple temporal scales. We also introduce a Behavior-Aware Modulation (BAM) block, inspired by SE-Networks, which modulates temporal embeddings to emphasize behavior-relevant features prior to attention. We evaluate on two datasets: RatSI (5 behavior classes, 12D pose inputs) and CalMS21 (4 behavior classes, 28D pose inputs). On RatSI, MSGL-Transformer achieves 75.4% mean accuracy and F1-score of 0.745 across nine cross-validation splits, outperforming TCN, LSTM, and Bi-LSTM. On CalMS21, it achieves 87.1% accuracy and F1-score of 0.8745, a +10.7% improvement over HSTWFormer, and outperforms ST-GCN, MS-G3D, CTR-GCN, and STGAT. The same architecture generalizes across both datasets with only input dimensionality and number of classes adjusted.

Figures & tables

Figure 1: Four consecutive frames for each of the five behaviors in the RatSI [ 29 ] dataset, showing how each behavior develops over time. Behavior instance durations vary across videos, with Solitary and Following typically spanning longer periods and Moving Away occurring over shorter spans.

Video	Solitary	Approaching	Following	Moving Away	Social Nose Contact	Allogrooming	Nape Attacking	Pinning	Other	Uncertain	Total Frames
Observation01	12,281	1,405	2,150	985	1,799	1,778	221	16	792	88	21,515
Observation02	14,709	1,066	593	592	1,754	2,240	95	0	231	55	21,335
Observation03	15,428	1,083	2,522	482	1,106	587	104	1,036	472	21	22,841
Observation04	14,776	1,528	444	931	2,621	2,003	141	0	300	0	22,744
Observation05	14,297	1,310	2,955	880	1,067	111	434	127	1,678	0	22,859
Observation06	13,152	2,443	2,141	1,503	1,980	486	192	0	856	0	22,753

Table 1: Frame distribution of social behaviors across all nine RatSI observation videos. The last row indicates the overall percentage of each behavior across all videos (total 202,550 frames).

Behavior	Description / Social Context
Solitary	The rat is alone , showing no social engagement (isolation phase).
Approaching	The rat initiates interaction by moving toward another (beginning of engagement).
Following	Indicates sustained interaction , maintaining social attention.
Moving Away	Shows withdrawal or avoidance , marking disengagement.
Social Nose Contact	Reflects direct physical interaction , the highest level of social engagement.

Table 2: Summary of the selected RatSI behaviors that are considered to represent the core range of rodent social interactions.

Figure 2: Sample frames from the RatSI dataset showing the five social behaviors studied in this work: Solitary, Approaching, Following, Moving Away, and Social Nose Contact, spanning from no social engagement to direct physical contact.

Behavior	Description / Social Context
Attack	Aggressive contact initiated by the resident mouse toward the intruder.
Investigation	The resident mouse closely sniffs or investigates the intruder mouse.
Mount	The resident mouse mounts the intruder, reflecting dominance behavior.
Other	All remaining frames not belonging to the three active social behaviors.

Table 3: Summary of CalMS21 behaviors representing the core range of mouse dyadic social interactions.

Figure 3: Representative video frames from the CalMS21 dataset illustrating the four annotated behavior classes. The resident (black) mouse initiates all active behaviors toward the intruder (white) mouse in the resident-intruder assay [ 43 ] .

Split	Attack	Investigation	Mount	Other	Total Frames
Train (70 videos)	14,039	146,615	28,615	318,469	507,738
Test (19 videos)	12,630	61,275	31,848	156,354	262,107
All Videos	26,669	207,890	60,463	474,823	769,845
Train (%)	2.76	28.88	5.64	62.73	—

Table 4: Frame distribution of behavior classes across training and test splits of the CalMS21 Task 1 dataset (total 769,845 frames across 89 videos).

Aspect	RatSI	CalMS21
Species	Rats	Mice
Number of Videos	9	89
Total Labeled Frames	~202,550	~769,845
Keypoints per Animal	3	7
Input Dimension ( $D$ )	12	28
Behavior Classes ( $C$ )	5	4

Table 5: Comparison of the two evaluation datasets.

Figure 4: Overview of the proposed MSGL-Transformer architecture. The model combines global token generation, behavior-aware modulation, and multi-scale attention within a lightweight transformer encoder to jointly capture global context and fine-grained temporal dynamics in rodent social interactions.

Figure 5: Structure of the multi-scale attention module. The modulated input $\hat{Z}_{0}$ is passed through three parallel branches: short-range causal attention on the first $\lfloor T/2\rfloor$ frames (this branch focuses on rapid motion cues occurring within short temporal contexts), medium-range causal attention on all $T$ frames, and global bidirectional attention on all $T+1$ tokens. The short-range and medium-range outputs are averaged on overlapping positions to produce $A_{local}$ , which is combined with $A_{global}$ through residual addition and layer normalization.

Hyperparameter	Symbol / Setting	Value	Description
Sequence length	$T$	35	Number of time steps per sequence
Input dimension	$D$	12 (RatSI) / 28 (CalMS21)	Coordinate features per frame
Output classes	$C$	5 (RatSI) / 4 (CalMS21)	Number of behavior classes
Model dimension	$d_{\text{model}}$	64	Embedding size for each token
Feed-forward dimension	$d_{\text{ff}}$	128	Hidden size of the FFN
Attention heads	$h$	4	Number of heads

Table 6: Summary of model and training hyperparameters used in the proposed MSGL-Transformer.

Window Size	Parameters	Accuracy	F1-Score	Investigation F1
$T=20$	223,492	0.8706	0.8751	0.7728
$T=35$	285,892	0.8709	0.8745	0.7881
$T=50$	348,292	0.8711	0.8744	0.7896

Table 7: Effect of window size on model performance on CalMS21 Task 1.

Validation	Test	Precision	Recall	F1	Accuracy
2	8	0.716	0.724	0.714	0.724
3	9	0.667	0.670	0.663	0.670
4	1	0.746	0.709	0.722	0.709
5	6	0.747	0.773	0.745	0.773
6	4	0.773	0.795	0.778	0.795
7	3	0.796	0.799	0.797	0.799

Table 8: Cross-validation results of the MSGL-Transformer on the RatSI dataset.

Figure 6: Accuracy, precision, recall, and F1-score across three representative splits in RatSI: Valid-7–Test-3 (best-performing), Valid-2–Test-8 (near the mean), and Valid-3–Test-9 (most challenging).

Class	Valid-2-Test-8	Valid-3-Test-9	Valid-7-Test-3
Approaching	0.557	0.595	0.453
Following	0.566	0.520	0.592
Moving away	0.142	0.194	0.073
Social Nose Contact	0.624	0.510	0.348
Solitary	0.863	0.798	0.909

Table 9: Per-class F1-scores across three representative splits in RatSI dataset

Figure 7: ROC curves across the three representative splits of the RatSI dataset.

Figure 8: Confusion matrices obtained for the Valid-2–Test-8, Valid-3–Test-9, and Valid-7–Test-3 splits.

Model	Accuracy	Precision	Recall	F1-score
TCN	0.7519	0.7457	0.7519	0.7393
LSTM	0.6929	0.8361	0.6929	0.7392
Bi-LSTM	0.6865	0.8414	0.6865	0.7377
MSGL-Transformer (Ours)	0.8148	0.7976	0.8148	0.7967

Table 10: Comparison with different baseline models on the RatSI dataset (Valid-8–Test-2 split).

Model Variant	BAM	MSA	Accuracy
Base	$\times$	$\times$	0.7567
Base + Multi-Scale Attention (MSA)	$\times$	✓	0.7798
Base + BAM Module	✓	$\times$	0.7859
Full Model (MSGL-Transformer)	✓	✓	0.8148

Table 11: Ablation study results on the Valid-8–Test-2 split of the RatSI dataset.

Class	Precision	Recall	F1-Score	Support
Attack	0.5942	0.5720	0.5829	12,586
Investigation	0.7157	0.8768	0.7881	60,964
Mount	0.8646	0.8798	0.8722	31,813
Other	0.9774	0.8909	0.9321	156,079

Table 12: Per-class results of the MSGL-Transformer on CalMS21 Task 1.

Figure 9: ROC curves of the MSGL-Transformer on the CalMS21 test set, showing per-class discrimination performance.

Figure 10: Confusion matrix of the MSGL-Transformer on the CalMS21 test set. We can see the model does best on the dominant Other class, while Attack is the hardest class. This reflects the severe class imbalance in the dataset.

Model	Accuracy	Precision	Recall	F1-Score
TCN	0.8486	0.8825	0.8486	0.8588
LSTM	0.8683	0.8714	0.8683	0.8696
Bi-LSTM	0.8606	0.8682	0.8606	0.8638
MSGL-Transformer (Ours)	0.8709	0.8842	0.8709	0.8745

Table 13: Baseline comparison on CalMS21 Task 1 using the official test set provided by Sun et al. [ 43 ] .

Model Variant	BAM	MSA	Accuracy	Attack F1
Base (no MSA, no BAM)	$\times$	$\times$	0.8771	0.5635
Base + MSA only	$\times$	✓	0.8677	0.5518
Base + BAM only	✓	$\times$	0.8786	0.5221
Full (MSA + BAM)	✓	✓	0.8709	0.5829

Table 14: Ablation study on CalMS21. The full model achieves the highest F1 for Attack, the most class-imbalanced behavior (2.76% of training frames).

Figure 11: Boundary error analysis results on the CalMS21 dataset. The left plot illustrates the variation in prediction accuracy with respect to the distance from the nearest behavior transition, while the right plot presents the distribution of samples observed at each distance.

Method	Input Type	Avg. Per-Class Acc (%)	Year
ST-GCN †	Keypoints	74.5	2018
MS-G3D †	Keypoints	73.5	2020
CTR-GCN †	Keypoints	71.6	2021
STGAT †	Keypoints	73.9	2021
HSTWFormer † [ 38 ]	Keypoints	76.4	2024
MSGL-Transformer (Ours)	Keypoints	87.1	N/A

Table 15: Comparison with published keypoint-based methods on CalMS21 Task 1. Results for methods marked with

\dagger

are reported from [ 38 ] .

Model	Parameters	Time/Epoch (s)
TCN	30,788	$\sim$ 97
LSTM	57,604	$\sim$ 87
Bi-LSTM	147,972	$\sim$ 104
MSGL-Transformer (Ours)	285,892	$\sim$ 295

Table 16: Computational complexity of models on CalMS21 using an NVIDIA RTX A4000 GPU.

Model	RatSI Acc	RatSI F1	CalMS21 Acc	CalMS21 F1
TCN	0.7519	0.7393	0.8486	0.8588
LSTM	0.6929	0.7392	0.8683	0.8696
Bi-LSTM	0.6865	0.7377	0.8606	0.8638
MSGL-Transformer (Ours)	0.8148	0.7967	0.8709	0.8745

Table 17: Performance results of all evaluated models on the RatSI and CalMS21 datasets. For the RatSI dataset, the Valid-8–Test-2 split was used for evaluation, while for the CalMS21 dataset the Task 1 test set provided in Sun et al. [ 43 ] was used.

Explore similar work

Rodent-Bench
Feb 20, 2026Thomas Heap, Laurence Aitchison, Emma Cahill +1Animal Behavior Analysis Video-Language Model Evaluation
University of Bristol, Bristol, UK
RatSeizure: A Benchmark and Saliency-Context Transformer for Rat Seizure Localization
Mar 25, 2026Ting Yu Tsai, An Yu, Lucy Lee +5Animal Behavior Analysis Temporal Action Localization
University at Albany, SUNY · University of Louisville · GE HealthCare
From Language to Behavior: Scaling Sequence Transformers for Industrial Recommendation Ranking with Rec-Native Designs
Sep 1, 2026Jie Chen, Xiangqian Yu, Yanchao Lian +9Sequential Recommendation Learning to Rank
ByteDance

Feb 20, 2026J/K move · Enter open · S save

Rodent-Bench

Thomas Heap, Laurence Aitchison, Emma Cahill, Adriana Casado Rodriguez

University of Bristol, Bristol, UK

Animal Behavior Analysis Video-Language Model Evaluation Multimodal Large Language Models Multimodal Model Evaluation

Open paper AskPDF arXiv

We present Rodent-Bench, a novel benchmark designed to evaluate the ability of Multimodal Large Language Models (MLLMs) to annotate rodent behaviour footage. We evaluate state-of-the-art MLLMs, including Gemini-2.5-Pro, Gemini-2.5-Flash and Qwen-VL-Max, using this benchmark and find that none of these models perform strongly enough to be used as an assistant for this task. Our benchmark encompasses diverse datasets spanning multiple behavioral paradigms including social interactions, grooming, scratching, and freezing behaviors, with videos ranging from 10 minutes to 35 minutes in length. We provide two benchmark versions to accommodate varying model capabilities and establish standardized evaluation metrics including second-wise accuracy, macro F1, mean average precision, mutual information, and Matthew's correlation coefficient. While some models show modest performance on certain datasets (notably grooming detection), overall results reveal significant challenges in temporal segmentation, handling extended video sequences, and distinguishing subtle behavioral states. Our analysis identifies key limitations in current MLLMs for scientific video annotation and provides insights for future model development. Rodent-Bench serves as a foundation for tracking progress toward reliable automated behavioral annotation in neuroscience research.

Abstract

Explore similar work

Rodent-Bench

RatSeizure: A Benchmark and Saliency-Context Transformer for Rat Seizure Localization

From Language to Behavior: Scaling Sequence Transformers for Industrial Recommendation Ranking with Rec-Native Designs