johnolafenwa / DeepStack

The World's Leading Cross Platform AI Engine for Edge Devices
Apache License 2.0
707 stars 113 forks source link

Yolo V4 #11

Closed robmarkcole closed 4 years ago

robmarkcole commented 4 years ago

Any thoughts about implementing this? https://github.com/sicara/tf2-yolov4

johnolafenwa commented 4 years ago

Hello @robmarkcole , thanks for sharing this. We are planning to switch to https://github.com/facebookresearch/detr as it is more accurate and can be more optimised to run faster than yolo. The object Detection api will be significantly improved after we integrate this both for the pre trained and custom models. I will make updates for the integration this week. Will be addressing most of the issues in the next few days.

By the way, will take some time to compare the two together, since DETR is anchorless and without need for NMS, it will be significantly easier to train on custom objects.

robmarkcole commented 4 years ago

sounds like a really interesting update, if I can help out let me know, been meaning to get my feet wet with pytorch & ecosystem

johnolafenwa commented 4 years ago

Great to know. Sure, i will reach back with more info on the structure of the existing code to make contributing easier.

johnolafenwa commented 4 years ago

Hello @robmarkcole , I just added the new object detection model. It is significantly more accurate on both day and night images while being fast. Check https://github.com/johnolafenwa/DeepStack/commit/34d70ac33abfd29f50c81587558174b055465c77

Build and test pipelines will soon be added to enable easy testing of this.

You can check the results here. https://github.com/johnolafenwa/DeepStack/tree/dev/results

Going further, new updates will be structured as PRs that we shall review together.