low fps and no bounding box on Tensorflow 2

[Sachin Mohan]

I was testing this repository https://github.com/susnato/keras-yolo3 by @susnato on Tensorflow 2.6.0

But when I run the below

python yolo_video.py --model model_data/yolo.h5  --input /home/mohan/git/backups/drive.mp4 --anchors model_data/yolo_anchors.txt --classes model_data/coco_classes.txt

I am getting low fps and no bounding boxes are generated too. Can someone please help.

I was getting the below error when running the inference before

ValueError: Subshape must have computed start >= end since stride is negative, but is 0 and 2 

So I applied the fix given here -> https://stackoverflow.com/questions/57558476/training-a-keras-model-yields-multiple-optimizer-errors

Now my output looks like the below.

low_fps_no_bounding_boxes_3-2022-04-25_20.22.42.mp4

[Susnato Dhar]

Hi @sachinkmohan , can you please provide some additional information about how to reproduce this issue? Any Colab notebook or something would be useful.
Actually I created TF 2.6 version of this repo a long time ago and don't remember the steps but will look for the cause of this issue and try to fix it.

[Sachin Mohan]

Thanks @susnato for writing back! :)

I made a fork of your repository -> https://github.com/sachinkmohan/keras-yolo3
Make sure you checkout the tf_2.x_compatible branch to see the updated README file and commits.

Once you have everything in place just execute the below. Please make sure you convert and generate yolo.h5 file as given in the old README.

python yolo_video.py --model model_data/yolo.h5  --input /home/mohan/git/backups/drive.mp4 --anchors model_data/yolo_anchors.txt --classes model_data/coco_classes.txt

[ksk]

Did anyone solve this issue? I have the same issue. The prediction results look like it correctly found objects, but every bounding box is located left-top of the image.

[EdgeNeko]

Same problem, looking for the latest update of this issue

[ksk]

I figured out the problem was due to the shape.