First of all, thank you for sharing this brilliant work. I want to build an on-road vehicle detector, hence I was trying out your code with MS coco dataset with resnet101. I used your pretrained weights named 'coco_900-1190k.tgz' to test on coco-val2017 dataset. But the results I got were a bit disappointing, all the required bounding boxes were there, but there were just a lot of bounding boxes which should have had been suppressed by NMS. I tried to reduce the IoU-threshold to even 0.05, but it didn't work out!
It feels like Non-Max Suppression probably ain't working. I am attaching a few files for your reference. Please look into it, if anything can be done.
Hello @endernewton ,
First of all, thank you for sharing this brilliant work. I want to build an on-road vehicle detector, hence I was trying out your code with MS coco dataset with resnet101. I used your pretrained weights named 'coco_900-1190k.tgz' to test on coco-val2017 dataset. But the results I got were a bit disappointing, all the required bounding boxes were there, but there were just a lot of bounding boxes which should have had been suppressed by NMS. I tried to reduce the IoU-threshold to even 0.05, but it didn't work out!
It feels like Non-Max Suppression probably ain't working. I am attaching a few files for your reference. Please look into it, if anything can be done.