DNN module fails to find some large objects

I have a YOLOv3-tiny network I trained using Darknet. I can successfully load that network using cv::dnn::readNetFromDarknet() and I can get the predictions. Everything mostly works.

But large objects somehow aren’t “seen”. Here is an example where you can see the objects detected are exactly the same, except for the pedestrian and the two cars. The image on the left is using Darknet, the image on the right is with OpenCV’s DNN module from OpenCV v4.5.3.

Source video with many more examples: Compare 4 YOLO results - YouTube

This is not an isolated example. All my tests suffer from the same thing, where objects in the foreground (often large objects) are undetected when I use OpenCV DNN vs Darknet.

Curious to know if someone might have an idea as to the cause, or what I need to fix.