Add state of the art YOLOX model from YOLO series.

[bogdan-evtushenko]

🚀 Feature

Is your feature request related to a problem? Please describe. The problem is basically in not sufficiency of mAP score for not big dataset size available.

Describe the solution you'd like Adding model the newest version from YOLO series - YOLOX - to the list of available models to use (github) (arxiv)

Describe alternatives you've considered We did a couple of experiments and that's what we had:

• using YOLOv5 - the performance was worse (as it supposed to be).
• using EfficientDet has shown a little bit worse performance and training time as well (At least with our experiments).

[roboserg]

Why YOLOX, and not YOLOR? YOLOR has better COCO AP over YOLOX - https://miro.medium.com/max/875/0*GHtVCZCRiHWhaQfI

YOLOX is even worse then Scaled-YOLOv4

[bogdan-evtushenko]

It's true, so from what I see from best to worse (COCO AP)(@roboserg correct me if I'm wrong): YOLOR -> Scaled-YOLOv4 -> YOLOX -> YOLOv5 So adding any of those could improve icevision usage experience.

[lgvaz]

@bogdan-evtushenko or @roboserg would you be up to adding this? It's a lot of work but should be straight forward, it should follow the same structure as our other models. If you're interested we can have a more detailed chat =)

[weiji14]

YOLOX is available in mmdetection (https://github.com/open-mmlab/mmdetection/tree/v2.16.0/configs/yolox), so might be easier to integrate into icevision at https://github.com/airctic/icevision/tree/master/icevision/models/mmdet/models :wink:

[bogdan-evtushenko]

@bogdan-evtushenko or @roboserg would you be up to adding this? It's a lot of work but should be straight forward, it should follow the same structure as our other models. If you're interested we can have a more detailed chat =)

Yeap, I'm interested in adding that, let's have a chat?

[lgvaz]

@bogdan-evtushenko yeeep, are you in the discord server already?

[FraPochetti]

This is done.