yolov4-tiny: coco14 or coco17?

[zhimengf]

In the paper of scaled-yolov4, the author said:

We use MSCOCO 2017 object detection dataset to verfy the proposed scaled-YOLOv4

So it means yolov4-tiny is trained on coco2017.

From the first author's repo: https://github.com/WongKinYiu/ScaledYOLOv4/blob/yolov4-large/data/coco.yaml

we can see he used train2017.txt for training and val2017.txt for validation.

So my question is is the yolov4-tiny.weights pretrained model here is also trained on train2017.txt rather than trainvalno5k.txt?

You know, they are two different splits.

[zhimengf]

I confirmed it is trained on coco17, as mAP on 17 validation is lower than 5k.txt on coco14 yolo split. Recall that yolo split and coco17 split has overlap

[zhimengf]

closing now