Spring 2025
Weeks 1 & 2 – Monday, January 6, 2025
Summary
Kailey & Charlie: Looking for potential conferences for BioBoost, changing format of paper to chosen publication, and begin major BioBoost rewrite.
Eric: Working on selecting a triplet mining strategy for Cichlid reID.
Recording Transcript: Weeks 1-2 Meeting Transcript.docx
Slides: Weeks 1-2 Meeting Slides
Week 3 – Monday, January 13, 2025
Summary
Kailey: Started writing up the WACV BioBoost paper for Ecological Informatics.
Charlie: Having computer problems, looking at ways to keep things running with new dataset, and getting PACE access
Eric: Proposed DAS + RHDIS + Distance Weighted Sampling strategies for triplet selection. Working on finding code for these strategies.
Recording: Week 3 Meeting Recording.mp4
Slides: Week 3 Meeting Slides
Week 4 – Monday, January 27, 2025
Summary
Kailey: Modified Bree’s SORT code to be able to run the modified SortFish on the new dataset. Modified Bree’s YOLO/SORT visualization code to work with the Lindenthal dataset. Investigated and fixed some problems in BioBoost pipeline.
Charlie: Worked on re-running SORT and cropping tracks into videos. Continued reviewing manuscript submission guidelines for BioBoost project.
Eric: Reviewed current version of BioBoost rewrite, and reviewed manuscript submission guidelines. Cleaned up the BioBoost GitHub by adding docstrings, type hints, and more guidance to README.
Recording: Cichlid CV Weekly Meeting-20250127_144424-Meeting Recording – View-only
Slides: Cichlid CV Week 4 Meeting Slides
Week 5 – Monday, February 3, 2025
Summary
Kailey: Looking into adding confidence labels to SORT detections and tweaking parameters to refine SORT output.
Charlie: Helping tweak the parameters of SORT to improve output. Uploading .bag files to Dropbox so we can test modifications to conversion scripts.
Eric: Adjusting the scripts that convert the ROS .bag files to .mp4 videos to fix speed variations in the footage. Adding Dockerfiles to these scripts to avoid needing a specialized OS to run the scripts.
Recording: Cichlid CV Weekly Meeting-20250203_143627-Meeting Recording – View-only
Slides: Week 5 Meeting Slides
Fall 2024
Week 6 – Wednesday, September 25, 2024
Meeting Summary
Repository Project: The team discussed pipeline issues, including resolving errors in the depth columns (marked as “Data Dedupe = yes”) and ensuring experiments are properly set up before using commands other than mc_multi. They strategized on dividing pipeline responsibilities among members to avoid overwhelming any single person, emphasizing caution to prevent data corruption on Dropbox. Additionally, the team agreed to check on changes being made by McGrath’s new student and highlighted the need for branch visualization showing code authorship, contact details, and active years for better collaboration.
DLC Project: The team located Adam’s DLC code at this repository and briefly addressed miscellaneous topics related to the project.
Week 7 – Wednesday, October 2, 2024
Summary
Jeanette: Monitored Pis and videos nightly, addressed confusion around the “Tank Reset Start/Stop” on the controller sheet, and created documentation for the “Controller sheet.”
Charlie: Compared PVT-v2-B2 and PyraT-CAiT Small models, suggesting steps to reduce overfitting in PyraT-CAiT Small.
Kailey: Transitioned to working on DLC, encountered feasibility issues with AnyDesk, and planned tasks including studying DLC and meeting with Bree for next steps.
Thuan: Identified and mitigated memory and GPU bugs in DeepLabCut training, raising issues and fixes to the DLC GitHub repository.
Ruiqing: Completed DLC training and evaluation, achieving a test error of 27.68 px with a ResNet-based network.
Slides: CichlidCV_Week07_TeamMeeting_Slides.pptx
Recording: Cichlid CV Weekly Team Meeting-20241001_205351-Meeting Recording.mp4
Week 8 – Wednesday, October 9, 2024
Summary
Jeanette: Monitored Pis for irregular behavior, added error handling for camera issues, and created a Planner board for tracking code changes.
Charlie: Focused on improving PyraT-CAiT for reID by prioritizing triplet loss over classification, exploring WildlifeReID-10k, and refining the triplet selection pipeline.
Kailey: Addressed various DeepLabCut (DLC) training errors, proposed solutions for performance optimization, and planned tasks for evaluation and conference preparation.
Thuan: Trained DLC on cichlid data, created annotated videos, documented training workflows on GitHub, and supported teammates with setup and troubleshooting.
Ruiqing: Installed DLC on PACE, started network training, refined datasets, and worked on improving model performance on new video tests.
Slides: CichlidCV_Week08_TeamMeeting_Slides.pptx
Recording: Cichlid CV Weekly Team Meeting-20241008_210237-Meeting Recording.mp4
Week 9 – Wednesday, October 16, 2024
Summary
Jeanette: Investigated Pis outages and began developing a notification system via Google Apps Scripts to automate alerts for issues.
Charlie: Proposed a triplet selection strategy and outlined a plan to optimize the PyraT-CAiT reID model using WildlifeReID-10K, aligning triplet and cross-entropy losses for better performance.
Kailey: Encountered technical issues with video labeling and model analysis, planning to adjust batch sizes and troubleshoot encoding while preparing for the conference.
Thuan: Advanced DeepLabCut testing on PACE, generating labeled videos with tracking effects, and updated the DLC guide on GitHub.
Ruiqing: Annotated 500 frames, trained a model, analyzed lizard movement, and focused on refining labels and collecting more data for validation.
Slides: CichlidCV_Week09_TeamMeeting_Slides.pptx
Recording: Cichlid CV Weekly Team Meeting-20241015_210341-Meeting Recording.mp4
Week 10 – Wednesday, October 23, 2024
Summary
Jeanette: Monitored Pis, fixed duplication issues, wrote an email notification script for errors, and planned to enhance pipeline documentation and error checks.
Charlie: Detailed triplet mining strategies and developed formulas for combining substrategies to optimize PyraT-CAiT performance based on model thresholds.
Kailey: Annotated videos, trained models, resolved setting overwrite issues, and planned additional runs to improve model performance while gaining familiarity with DLC.
Ruiqing: Processed and organized large video datasets, planned to refine label sizes, and aimed to explore new tools like SLEAP for video analysis.
Thuan: Studied Adam’s ViT re-identification code for integration with DeepLabCut 3.0, planning to consult Adam for a deeper understanding and improve inference pipeline compatibility.
Slides: CichlidCV_Week10_TeamMeeting_Slides.pptx
Recording: Cichlid CV Weekly Team Meeting-20241022_210327-Meeting Recording.mp4
Week 11 – Wednesday, October 30, 2024
Summary
Jeanette: Monitored Pis and addressed duplication errors, created and tested an email notification script for errors, and planned to improve documentation and error handling in the pipeline.
Charlie: Explored methods to handle embedding size discrepancies in triplet mining, focusing on linear transformations and initializing key components of the training pipeline for PyraT-CAiT.
Kailey: Received BioBoost feedback, worked on improving DLC model test RMSE, and investigated snapshot evaluation issues while training additional models to resolve overfitting concerns.
Ruiqing: Processed and labeled large video datasets, reran network training for improvement, and began exploring the use of SLEAP software for future annotation tasks.
Thuan: Investigated implementing transformer re-identification for DeepLabCut 3.0, identified missing functionality in the PyTorch engine, and planned to extract and format features required for triplet dataset creation and transformer training.
Slides: CichlidCV_Week11_TeamMeeting_Slides.pptx
Recording: Cichlid CV Weekly Team Meeting-20241029_205955-Meeting Recording.mp4
Week 12 – Wednesday, November 6, 2024
Summary
Jeanette: Monitored Pis, identified recurring HTTP 401 errors causing outages, and planned to address these errors to prevent nightly disruptions.
Charlie: Focused on debugging and testing the DataDistiller and PyraT-CAiT model, planning next steps to refine its architecture for end-to-end reID and improve subloss performance.
Kailey: Evaluated DLC model snapshots, identified issues with PyTorch evaluation runtimes and inconsistencies, and worked on multiple projects including BioBoost and Lizard with plans to assist with DLC and reID tasks.
Ruiqing: Used SLEAP for data annotations, transferred labeled videos, and created visualizations for further analysis.
Thuan: Successfully implemented and ran Adam’s DeepLabCut re-identification workflow for the PyTorch engine, planning to evaluate transformer re-ID pipelines on improved triplet datasets with better video selections.
Slides: CichlidCV_Week12_TeamMeeting_Slides.pptx
Recording: Cichlid CV Weekly Team Meeting-20241105_210828-Meeting Recording.mp4
Week 13 – Wednesday, November 13, 2024
Summary
Jeanette: Monitored Pis, resolved campus power outage issues, and aimed to address a recurring error with Pi 25 while exploring improvements to error logging and automated restarts.
Charlie: Concluded initial debugging of DataDistiller, identified loss imbalance in PyraT-CAiT reID, and planned to integrate pre-trained models into the DLC reID pipeline as part of ongoing collaborative projects.
Kailey: Progressed on BioBoost, DLC + ReID, and Lizard projects, testing SimBA with cichlid videos and preparing to evaluate Thuan’s PyTorch reID implementation.
Ruiqing: Resolved data allocation issues, continued working with SLEAP for network creation, and focused on data visualization and validation.
Thuan: Updated a guide for training fish reID on DeepLabCut’s PyTorch engine, studied TransReID’s potential for fish datasets, and assessed its applicability given limitations in temporal and viewpoint data.
Slides: CichlidCV_Week13_TeamMeeting_Slides.pptx
Recording: Cichlid CV Weekly Team Meeting-20241112_210555-Meeting Recording.mp4
Week 14 – Wednesday, November 20, 2024
Summary
Jeanette: Monitored Pis, addressed recurring errors, and planned to implement automatic restart functionality and enhanced error logging for Pis.
Charlie: Developed a custom COCO-to-YOLO conversion notebook, trained YOLOv5 on converted data, and outlined next steps including video annotation and pipeline integration for BioBoost.
Kailey: Explored B-SOiD for behavior clustering in the Lizard project, noting its advantages over SimBA, and progressed on tasks related to BioBoost, DLC + ReID, and cichlid datasets.
Thuan: Studied TransReID and tested DeepLabCut re-ID implementations, integrating ViT embeddings but encountered memory issues requiring further troubleshooting and consultation with Adam.
Ruiqing: Continued SLEAP network training, relabeled datasets, and addressed technical issues with TensorFlow while preparing to analyze lizard data further.
Week 15 – Wednesday, November 27, 2024
Summary
Jeanette: Developed and tested a system using Pi 47 to automatically reset Pis encountering errors and planned to log errors in CSV files alongside existing email notifications.
Charlie: Successfully automated video extraction from ROS 1 .bag files using a Python script and prepared for large-scale video processing after resolving storage issues.
Kailey: Progressed on the Lizard project by exploring B-SOiD and other behavior analysis tools, creating documentation for Dr. Stroud, and planned to refine clustering and expand tool comparisons.
Thuan: Resolved memory errors in Adam’s re-ID pipeline, improved accuracy to 95-100%, and identified and raised a bug fix for DeepLabCut’s transformer re-ID code, significantly reducing training time.
Ruiqing: Continued SLEAP training to address overfitting, verified labeled videos on Dropbox, and worked on finalizing reports for DLC and SLEAP.
Week 16 – Wednesday, December 4, 2024
Summary
Jeanette: Continued monitoring Pis and collaborating with the lab liaison to maintain functionality and address any issues.
Charlie: Fixed a bug in video extraction from .bag files, enabling successful playback of IR video data, and planned to process all videos through the YOLOv5s pipeline and SORT.
Thuan: Documented the logic and implementation of model accuracy calculations in DeepLabCut’s triplet processor and planned to create detailed documentation for reproducibility.
Kailey: Worked on egocentric alignment and latent space embedding for behavioral analysis using VAME, and ran B-SOiD on the cichlid dataset, generating promising results despite data challenges.
Ruiqing: Continued experiments with new tools and collaborated with Kailey on behavioral analysis projects.
Slides: CichlidCV_Week16_TeamMeeting_Slides.pptx
Recording: Cichlid CV Weekly Team Meeting-20241203_212234-Meeting Recording.mp4