Spring 2025 Lizard Jaw Segmentation Group Week 2 Computational Meeting
Week 14 Group Meeting
Group Meeting Quick Summary: The meeting on November 20, 2024, focused on Jacob Dallaire’s lizard classification model, which uses MobileNet V2 for transfer learning but faces challenges due to dataset imbalances and misclassification, particularly with the Equestrians, resulting in 35% overall accuracy. Suggestions included improving documentation, adding visual data examples, exploring alternative models like BIOCLIP, and refining training with expert feedback. Updates also covered team projects, with Breanna Shi addressing next semester’s assignments and documentation needs. Philip Woolley’s project on model performance and Ayush Parikh and Mercedes Quintana’s work on image-processing workflows and landmarking were discussed, along with the importance of a literature review on landmarking challenges. Action items included refining workflows, improving dataset quality, and preparing for the next semester’s work and collaborations.
Group Meeting Recording: https://drive.google.com/file/d/1F-sRtig3s-jFl47TsEYCdrT6v_p-hC0B/view?usp=drive_link
Week 10 Group Meeting
Group Meeting Quick Summary: The discussions focus on improving an object detection project and addressing issues with remote lab access. In the object detection project, Jacob Dallaire and James Stroud discuss increasing the dataset size by annotating an additional 1,000 images and applying a 90° rotation to the existing training data for augmentation. This would quadruple the dataset and potentially improve the model’s performance. However, the model is currently underperforming, predicting only one class with 40% accuracy. They also note the varying quality of images, from good to bad, impacting model training. In the XRAY team discussion, there are challenges with the annotation tool’s website, leading to the sharing of a GitHub link as an alternative. Some team members have not yet used lab computers due to setup issues, with Mercedes Quintana planning to resolve these and start using the lab once technical problems are fixed.
Group Meeting Recording: https://drive.google.com/file/d/1w0vlLLa9It-3nAKBvKLQJX96FlFCfSK1/view?usp=drive_link
Week 8 Group Meeting
Group Meeting Quick Summary: Jacob Dallaire is refining a lizard detection model using transfer learning with MobileNet V2 and annotating 1,000 images to improve accuracy. Despite challenges like background complexity and small lizard sizes, he aims to enhance model performance by using LabelML for bounding box annotations and cropping images to create a more refined dataset. The team is awaiting the next model iteration to evaluate improvements. Philip Woolley discussed the time-consuming process of manually segmenting lizard head scans, resulting in 270 training and 75 validation images. He plans to continue segmentation and retrain the model with new data while implementing quality calculations to better assess training progress. The team also discussed image processing pipelines, debating manual vs. automated approaches. Ayush Parikh created an automated script for consistent adjustments like contrast, and Dr. Stroud raised the possibility of exploring deep learning for individualized image processing in the future.
Group Meeting Recording: https://drive.google.com/file/d/1etyitMH1NjDaTXrDNTI6IvrpBX4ORHS7/view?usp=drive_link
Week 6 Group Meeting
Group Meeting Quick Summary: The team discussed the progress of various machine learning and image processing projects. Jacob Dallaire is working on a species classification model using transfer learning with MobileNetV2, achieving near-perfect accuracy on training data, but facing issues with data splitting. He plans to retrain the model with corrected data and use a new GPU to speed up training. The team also discussed improvements in X-ray image processing for species landmark detection, with guidance from Dr. Porto leading to reduced errors in bounding box placement. Despite progress, challenges remain with extremities like toes and fingertips, and the dataset will be expanded for better model generalization. In the lizard jaw segmentation project, Philip Woolley is using a mask segmentation transformer model to automate the CT scan process. While initial results are promising, more data is needed to improve accuracy. The team aims to automate processing for comparative species analysis and integrate it for broader use.
Group Meeting Recording: https://drive.google.com/file/d/1IYJkIgffpBdWYI_gl9oLpGYZXWwYBZU7/view?usp=drive_link
Week 4 Group Meeting
Group Meeting Quick Summary: The meetings discussed ongoing projects in machine learning and image processing, focusing on X-ray datasets, species classification, and segmentation tasks. For the X-ray project, the team addressed discrepancies between manually and automatically processed image counts, highlighting the need for more data in the automatically processed set to enhance model training. The Anole Classifier project dealt with challenges of data imbalance and poor image quality, with plans to use oversampling, undersampling, and data transformations to improve classification. Meanwhile, the Deep Lab Card project made progress in video frame extraction and labeling, while a high-spec lab computer with an i9 processor and GTX 4090 was ordered for remote access and computing tasks. The Lizard Jaw Segmentation project faced issues with data alignment and segmentation accuracy but proposed a 3-step process—rotation, cropping, and machine learning—while exploring the use of pre-existing dental segmentation models. Action items include expanding datasets and refining segmentation methods.
Group Meeting Recording: https://drive.google.com/file/d/1YkvmCvUZRAV4K-B5pJ-KRLerezqGMf9Y/view?usp=drive_link
Week 2 Group Meeting
Group Meeting Quick Summary: In the XRAY team meeting, the focus was on visualizing project data and improving the presentation of technical results. The team discussed using CSV files to represent data on known and automated landmarks, with a color-coding scheme of red for manual landmarks and green for predicted ones, connected by lines to highlight alignment or discrepancies. The importance of presenting these technical findings in an accessible and digestible format for broader audiences was emphasized. Team members were encouraged to prepare a presentation for their respective projects, ensuring the results are communicated effectively. To enhance collaboration, James T. Stroud proposed setting up Discord channels for each project, where resources and updates could be shared easily. Action items included creating visual plots from CSV data and preparing presentations for upcoming meetings. These steps aim to improve both the internal workflow and external communication of project outcomes.
Group Meeting Recording: https://drive.google.com/file/d/1cfxKd77EMom5Cok2DrQlTkw7GetDZ68o/view?usp=drive_link