2 pointsby jmalevez6 hours ago1 comment
  • jmalevez6 hours ago
    I’ve spent a lot of time dealing with the class imbalance problem in computer vision where you have a lot of OK samples but not enough examples of the defects that actually matter.

    I built Silera to bypass the months spent hunting for and labeling these rare edge cases in the field.

    How it works: You upload one or a few reference images of the defect you care about (a scratch, a crack,...), OK samples and it generates thousands of unique, segmented variations in a couple of minutes. The goal is to produce a large amount of variation of the defect you care about to replace the time you'd spend collecting and labelling it in real life.

    Seeking Technical Feedback: I just launched the beta and I'm looking for CV engineers to stress-test the output.

    Specifically, I'm curious about:

    - Recall: Does this actually help you catch rare edge cases better than standard augmentations?

    - Precision: Are the masks clean enough for your IoU requirements, or is there still too much "leakage" on complex textures?

    It exports directly to COCO format for immediate training.

    I've added 100 free credits to new accounts for benchmarking.

    I’ll be around to answer any technical questions.

    Jérôme