ReCaptchav2

anyone knows how CapMonster and XEvil are able to beat recaptcha image challenge? What is their success rate? Are they using neural networks? If so is that supervised learning? Are they using datasets gathered from recaptcha itself or some public one like ImageNet or maybe both?

I experimented with various prediction/detection networks like ResNet, DenseNet, RetinaNet and my biggest challenge is to learn how to train them properly on custom data. For example I tried to manually categorize 1200 images of recaptcha buses and got very mixed results (some amazing detections but also lot of false positive ones)