medicaldetectiontoolkit VS albumentations

Checkout https://github.com/MIC-DKFZ/medicaldetectiontoolkit, they have a recent 3d rcnn implementation. Their paper is a good start on the topic. Also have a look at nnDetection from the same group, it might provide some more leads. Hth

I'm doing research in the area of biomedical image processing using deep learning. At the moment, I'm focused in the detection of biological structures in 3D microscopy images. For this task I trained a Faster R-CNN architecture, however I noticed that while the training loss is decreasing the validation loss is high, doesn't decrease and sometimes oscillates. When I looked at the results, the bounding boxes in the training and validation sets were too small but located at random spots. At first I thought that I should adjust the size of the anchors, however, even after trying different anchor sizes the results didn't improve. I would like to know if anyone has experience in working with 3D Faster R-CNN that could help me with suggestions for this project. (btw, I'm using the code available in this repo). Thanks in advance.

You can use albumentations if you are comfortable with using open source libraries https://github.com/albumentations-team/albumentations

One of the members of the core team of our open-source library https://albumentations.ai/

It was not the only reason he was hired; it was a solid addition to his already good performance at the interviews.

Or at least that is what the hiring manager later said.

I am one of the creators and maintainers of https://albumentations.ai/.

- 12800+ stars

Is something planned to support data augmentations? Something like https://albumentations.ai/

Here you go: https://albumentations.ai/

I think they should use some data augmentation techniques like I am using for Infinity AI if you wanna see more here. Note that most of these do not work for image generation.

Our second approach was a more automated one. Here the idea was to try out an image augmentation library, Albumentations, and use it for adversarial attacks. This time, instead of one-shot images, we applied the transformations at random time ranges. We chose for these transformations also to be more subtle than then one-shot images, such as vertical flips, grayscaling, downscaling, …

Check out albumentations. If your use case is segmentation, check out the offline augmentation of this project

I think this is the most popular open source project: https://github.com/albumentations-team/albumentations

You can also curate your own dataset by using open source images (https://universe.roboflow.com/search?q=faces%20images%3E1000) and open source augmentations (https://github.com/albumentations-team/albumentations). Or you can do use the augmentation UI (https://docs.roboflow.com/image-transformations/image-augmentation) to apply noise, blurring, shear, crop, etc.