Ideas for evals - Githubissues

royerloic commented 8 months ago

List of tasks, these were initially drawn for Omega, but can be adapted as functions for the purpose of this work:

[x] Convert RGB image to gray scale with configurable weights see
[x] Upscale and downscale image in arbitrary dimensions with different interpolation modes similar example
[ ] #6
[x] Do a max-min projection of a stack along an arbitrary axis for a nD image see and
[ ] #7
[x] Rotate 2D image around an arbitrary point (r_x, r_y) by an angle alpha. similar example
[ ] Gamma correction to the image.
[ ] Stitch two 2D images along a border using cross-correlation.
[ ] Sigmoidal threshold function with threshold and sharpness parameter
[ ] #37
[x] Gaussian filtering with sigma parameter. similar example
[ ] Box filtering of 2D image with box’s width and height parameters.
[ ] Butterworth filtering of image with cut off frequency parameter and order.
[x] Sharpening filter similar example
[x] Sobel edge filter similar example
[x] Lucy Richardson deconvolution that takes image and psf as input (two diff images) and the number of iterations. less specific example
[x] Convolve two images see
[ ] Rotate an image around its center by an angle alpha using the Fourier transform.
[ ] Return the FFT spectrum of an image as the absolute logarithm of the Fourier transform magnitude. Ensure that the DC component is at the center of the image. Use reflection padding and apodisation to reduce artefacts due to discontinuities at the image borders.
[ ] Arbitrarily oriented slice through a 3D image. Expose parameters for the normal to the slice plane and for the distance to the center of the image.
[ ] #8
[ ] Add noise to an image, first by applying a Poisson noise model and then a gaussian noise model. Expose the relevant parameters for both models.
[ ] Geometric mean filter for denoising an image with an arbitrary number of dimensions.
[ ] Alpha-trimmed mean filter to denoise a 2D image, expose relevant parameters if needed.
[ ] Adaptive median filtering of a 2D image.
[ ] Removal of periodic noise using notch filtering
[ ] Wiener filtering of an image
[ ] Rotate the hue of a RGB image with parameter in degrees from 0 to 360.
[ ] Convert gray-level image to RGB by applying a given LUT. LUT name and intensity min and max are parameters as well as gamma correction.
[ ] Segmentation of RGB image using k-means of pixel RGB vectors
[ ] Fusion of two 2D grayscale images using the Haar transform. Expose relevant/interesting parameters.
[ ] Fuse two gray-level images in the gradient domain.
[ ] Encode an invisible watermark in a gray-level image using DCT.
[ ] Find the best approximation of a 2D image as the sum of n Gaussians with different centers, sigmas, and weights.
[ ] Given an image, compute the local entropy map that for each pixel of the resulting image assigns the entropy calculated for the local patch.
[ ] Given a 3D array that represents 2D images over time, compute a 2 channel 2D image that encodes the x and y components of the PIV flow vectors for consecutive frames.
[ ] Make a widget that can apply n dilations and m dilations to a gray level image, add a parameter to control the order to apply these two gray morphology operators.
[ ] Make a widget that can apply hole filling to an arbitrary nD image. Expose relevant parameters.
[x] [nuclei image] Widget that segments nuclei on an image using Otsu thresholding and returns a labels layer. Expose relevant parameters, if appropriate. similar example
[ ] [image with lines] Widget that detects lines in a 2D gray level image.
[ ] Segment an RGB 2D image using the SLIC super pixel algorithm.
[ ] Segment a gray scale 3D image using the idea behind the SLIC super pixel algorithm.
[x] Make a widget that performs watershed segmentation on a 2D gray level image. The widget takes as input: (1) an image layer to segment, and (2) a points layer as input for markers. Expose other relevant parameters. similar example
[x] #9
[ ] Make a widget that computes the binary skeleton of a 3D image.
[ ] Make a widget that computes a median filter of variable radius. the code should be optimized with numba
[ ] Make a widget that finds circles or circle-looking structures in a 2D gray level image. First compute the edge image and then locate circles. The resulting image should be such that for each circle identified, the pixel corresponding to the center of the circle is assigned a value that is proportional to the radius. Variation: can you modify this function so that it draws the circles as disks in the resulting image? now the intensity of the drawn disks is proportional to how strong/liekly is the circle in the input image
[ ] Make a widget that given two 2D images, where the second is typically smaller than the first, finds occurrences of the second image by correlation. Return the first image but mask all pixels that are not covered by an occurrence of the second image.
[ ] Write a widget that takes an image layer and a labels layer and uses the catboost library to predict from all 5x5 overlapping patch of the input image, the class in the labels layer for the corresponding pixels. When learning to classify, only consider non-zero labels. Return the labels layer obtained by applying the classifier on the input image.
[ ] Write a widget that best approximates a 2D gray-level image as the sum of gray-level disks. Each disk has a position, radius, and pixel value, which need to be found to optimally approximate the image. Expose as a parameter the number of disks. Return the best-approximated image.

royerloic commented 8 months ago

More ideas here: https://github.com/royerlab/napari-chatgpt/blob/17644864cd4c343368f574844ce1ec8cdcb5497c/manuscript/SuppTable1_Example_widgets.pdf

haesleinhuepf commented 8 months ago

Great ideas, big thanks @royerloic ! I took the opportunity to turn your list into a checkbox-list. Some are ticked already because I implemented similar use-cases.

haesleinhuepf commented 8 months ago

A complete list of implemented test-cases can be found here: https://github.com/haesleinhuepf/human-eval-bia/blob/main/test_cases/readme.md

(and this list is updated semi-automatically)

haesleinhuepf / human-eval-bia

Ideas for evals #1