working on setting up obj detection for EfficientDet (latest detector from google, source released 3 days ago).
It has 3 heads (regressor, etc) so trying to figure out how to tie in these to FastAI - will post if I get it done today but please post if you’ve already done this?
I’m looking @muellerzr obj detection as guide from retinanet but there are some diffs.
Thanks!
3 Likes
Got it running with pure pytorch as that was faster. Accuracy is impressive.
Will try to get it into fastai tomorrow
7 Likes
Hi LessW2020, how is it going? Can you share the notebook when you finish?
Thanks.
Hi @LinHan - I am now working with EfficientDet in TF as they released their official code last week (and I have to be able to deploy to mobile) but yes I’ll post out a notebook once I get it working in FastAI.
4 Likes
Hi @ttsantos and all,
I switched to the DETR architecture which dispenses with the past 5 years of obj detection architecture and replaces it with a transformer.
For my work at least it blows everything else away bc detr understands relationships and global context…
Im trying to make a colab now so people can more easily train DETR for their own datasets but may be another week or two.its on my github if anyone is interested. https://github.com/lessw2020/training-detr
Im working with DETR full time though as using it for work production system.
Heres the link to their paper
https://arxiv.org/abs/2005.12872v3
I may have to use effdet for tflite mobile but for server detr is the new accuracy standard for my data.
12 Likes