Hacker News Reader

To dig deeper, there are couple of interesting nuances there:

- extreme number of objects on single photo, typical number of visible pieces in large pile is 1500-2000

- extreme number of classes in multi-class classification, there are ~1000 most common Lego bricks and up to 30000 classes if you include rare bricks and different patterns

- really hard data labelling: one photo can take up to a 5 work days to label

View on HN

I am really curious about the data labelling. Do you mean that you took photos of piles of Legos and labeled them by hand. Or individual Legos from different angles?

Also you mentioned data synthesis. How would this be possible? Unless your suggesting that you rendered photo realistic piles of Legos and used trained on them because if that is the case, please do a write up of the project. I can't imagine more interesting way to generate training data.

Very cool! How do you deal with occlusion? What kind of NN are you using? Would an off-the-shelf model like Yolo/SSD/FasterRCNN work for this or did you have to devise some new tricks?

Reply on news.ycombinator.com