Hagbard Celine
n3wb
I had this issue with cats and magpies being detected as person when using Standard Object Detection or IPcam-combined.
While browsing this forum, I saw several examples of people with similar issues. And no apparent solutions, other than train your own model.
I've been planning on trying to train a model, but I'm not there yet.
But I found a stop-gap solution and wanted to post it here. What I found is that if I replace Standard-Object-Detection/IPcam-combined with MegaDetector v5 , almost all these false positives disappear.
The catch is that this model is way bigger that the models supplied with CPAI and thus much longer inference time; around 250ms on a 3060Ti and around 900ms on a Xeon W-2255(both numbers with YOLOv5 6.2).
Also the model has only 3 possible results: animal(includes birds), person and vehicle(includes car, truck, train, boat and about everything in-between).
Also notable is that its trained on and reads pictures of 1280x1280 as opposed to 640x640 for the CPAI-included custom models.
While browsing this forum, I saw several examples of people with similar issues. And no apparent solutions, other than train your own model.
I've been planning on trying to train a model, but I'm not there yet.
But I found a stop-gap solution and wanted to post it here. What I found is that if I replace Standard-Object-Detection/IPcam-combined with MegaDetector v5 , almost all these false positives disappear.
The catch is that this model is way bigger that the models supplied with CPAI and thus much longer inference time; around 250ms on a 3060Ti and around 900ms on a Xeon W-2255(both numbers with YOLOv5 6.2).
Also the model has only 3 possible results: animal(includes birds), person and vehicle(includes car, truck, train, boat and about everything in-between).
Also notable is that its trained on and reads pictures of 1280x1280 as opposed to 640x640 for the CPAI-included custom models.
Last edited: