วีดีโอ

A challenging Cellpose problem from the Image.sc forum. In these images there are larger cells with very long protrusions and we have to tweak Cellpose training and evaluation parameters to try and optimize segmentation.

This video, though slightly rough around the edges, aims to demonstrate the performance of two deep learning approaches on bee data. The first approach is MobileSam, a combination of YOLO and SAM, and the second is Stardist with sparse labels. Starting from version 0.9, Stardist no longer requires complete rectangular ROIs for labeling; you can leave portions of the label empty. This feature ca...

This video shows how to use Napari (with a customized labeling widget) and PyTorch to train a deep learning UNET semantic segmentation model network to detect roots. A similar workflow could be used for image with vessels, cracks or other filament like structure. We use augmentation to supplement a very small number of labels.

This video shows how to use Stardist 2D to create a training set for Stardist 3D. We first create a simulated 3D image, then add blur, noise and random mis-alignment. Then we apply a stardist2D-connectin3D strategy to generate 3D labels. We train Stardist with these imperfect labels and show that Stardist3d does not overfit the imperfections.

A video showing how to create simulated images to test the performance of a custom Stardist model on objects at varying scale and shape (with some objects being very elongated).

In this video I show how to retrain a model after we get new data (in this case a new bee image). I talk about augmentation, test-prediction, validation-prediction, and self-prediction. Throughout the video I use Napari as both a viewer and a labeling tool.

Quick demo showing how to use the napari-segment-everything plugin. This is a plugin that can be used to explore the large, overlapping label collection that is returned from the Meta AI SAM segmentation model. The plugin can be installed from the Napari hub.

Short video demonstrating how to label an image in Napari with a combination of SAM and Napari labelling tools.

Shows how to quickly correct labels with Napari, then retrain a stardist model, in the case where another method is about 95% accurate but needs a few fixes.

This video was made to help with this imagesc image forum question forum.image.sc/t/how-to-count-bees-pattern-recognition-and-segmentation/90115/8 It shows how to label, train a stardist network for segmentation then relabel the bee image using ipython notebooks and the napari viewer. Code for the example is here github.com/True-North-Intelligent-Algorithms/tnia-python/tree/main/notebooks/image...

Deconvolution presentation for I2K 2022

How to do the IPython notebook part of assignment 3