Data required to have good results

[joansaurina]

Hey. In order to train a model I require to generate synthetic data.

You mention that if I run the blenderproc script 5 times, each time generating 1000 frames. Each frame will have five copies of the object and ten randomly chosen distractor object:

./run_blenderproc_datagen.py --nb_runs 1 --nb_frames 10 --path_single_obj ../models/Ketchup/google_16k/textured.obj --nb_objects 5 --distractors_folder ~/data/google_scanned_models/ --nb_distractors 10 --backgrounds_folder ../dome_hdri_haven/

How many frames are needed to train a good model for an specific object? More or less.

Thanks;

Joan

[TontonTremblay]

I think 5000 is a good start, I have seen decent results, but using 20k is probably better.

[sriram487]

@TontonTremblay I am trying to estimate pose of a single object. Following similar approach like yours. But instead of computing the belief map, trying to make the network regress the 2D projected vertices (in pixel co-ordinates) and later use the PnP to compute pose. Right now, using 6000 images for training and the model converges for the training data. I am also getting decent results in the test as well but not all the predictions are accurate. Do you think increasing the training sample will improve the performance of the model?

[TontonTremblay]

you doing like image -> (x,y) in normalized space? I would think that more data cannot go wrong.

[joansaurina]

And how many different background images?

[TontonTremblay]

I have ~2000 hdri background to download from. I used all of them. https://drive.google.com/file/d/1lp36MgTlS4OFaH0vdsTFhyGFJpQDY2YX/view?usp=drive_link you can download with the link.