DAMM

We introduce DAMM

Present deep learning tools for animal localization require extensive laborious annotation and time-consuming training for the creation of setup-specific models, slowing scientific progress. Additionally, the effectiveness of these tools in naturalistic settings is impeded by visual variability of objects and environmental diversity, hindering animal detection in complex environments. Our study presents the ’Detect Any Mouse Model’ (DAMM), a robustly validated object detector designed for localizing mice in complex environments. DAMM excels in generalization, robustly performing with zero to minimal additional training on previously unseen setups and multi-animal scenarios. Its integration with the SORT algorithm permits robust tracking, competitively performing with keypoint-estimation-based tools. These developments, along with our dissemination of DAMM, mark a significant step forward in streamlining ethologically-relevant animal behavioral studies.

Main Results

We train LVIS Mask R-CNNs on a high quality labortory mice detection dataset

We utilize Segment Anything Model (SAM) to speed up dataset annotation (left), and evaluate model preformance as more data is annotated (middle). We show test predictions fron the final model (right).

DAMM Generalizes Beyond its training

We demonstrate exceptional zero-shot detection results on a variety of unseen exparamental setups recorded in other labs, collected through OpenBehavior.

DAMM is robust to challenging conditions

DAMM is robust to challenging viewing angles, enviornemntal setups, mice fur coats, and recording quality (zero-shot predictions)

DAMM along with the SORT tracking algorithm enables robust tracking

DAMM enables video tracking on a variety of scenarios, both out of box and with additional data for fine tuning (click button for demo)

BibTeX

@article{kaul2024damm,
      author    = {Gaurav Kaul and Jonathan McDevitt and Justin Johnson and Ada Eban-Rothschild},
      title     = {DAMM for the detection and tracking of multiple animals within complex social and environmental settings},
      journal   = {bioRxiv},
      year      = {2024}
    }

DAMM: Detect Any Mouse Model

DAMM tracks mice zero-shot across diverse experimental setups