Roadmap

[lucasjinreal]

yolov7-d2 is keeping involving. We have made some progress on exploring more light-weights model while still able get a favorable mAP. For example, we have made a MobileOne-S0 version model based on YOLOX-Lite achieves mAP 29.5, exceed YOLOX-Nano with faster speed on CPU.

We also achieves mAP 44 with a Convnexttiny model based on YOLOX arch.

But, we are still trying to expanding the model matrix and exploring more efficient && effective models, If you are interested in training or adding new model inside yolov7-d2, it's more than welcome!.

Roadmap of Next Half-Year 2022

CPU realtime Instance segmentation

this would need several steps:

• A fast and efficient detector as base;
• An effective way to predict Masks;
• Boost the mAP and accuracy, at least better than SOLOv2;

CPU realtime Pose estimation

• A baseline version;
• Chasing KAPAO's result, get more accurate result

Replicate DINO's mAP

• Replicate DINO's mAP, reconstruct whole code base in DINO;
• Construct a more light-weighted transformer based model;