Nice visualizations!

[glenn-jocher]

The grid visualizations could be very useful for development and testing, and if fact YOLOv5 uses a new multiple-grid assignment structure that shows up well in your images.

If you have a concise plotting function for creating these perhaps we could do a PR into yolov5/utils/general.py (where all the plotting functions are located).

[seanavery]

@glenn-jocher Thanks dude!

Yeah I think it could be useful, I finally was able to understood the model output when I did the grid visualizations.

That sounds great. I still need to clean up the code some more, but will reach out with a PR this week.

[glenn-jocher]

@seanavery great, sounds good! Introspection is very important to help understand what is going.

[PonyPC]

Best works!

[seanavery]

@glenn-jocher hey, just a little update. getting back to the yolo code

Forked off from your repository and working on implementing a plotting function in general.py called plot_object_probability

One consideration, I am sending the pred output before non_max_suppression is applied in order to maintain the raw output from forward pass.

Example Ops:
1. input (1, 25200, 85)
2. -> [(1, 3, nx0, ny0, 85), (1, 3, nx1, ny1, 85), (1, 3, nx2, ny2, 85)] deconstruct the stacked output
3. -> [(1, nx0, ny0, 1), (1, nx1, nx2, 1), (1, nx2, ny2, 1)] take max object probability of the anchor box sizes per grid cell 
4. -> draw transparent rectangles in a threshold or hue range of the probability score

I was thinking it would be cool if there was a model "debug" mode where visualizations like the object probability, class grid, raw bounding boxes, and anchor statistic could be processed.

Making sure to use matplot lib to write images to inference/output directory where they can easily be displayed in the notebook.

[seanavery]

Object probability grid refactored to pytorch and working in the codebase. works on cpu and gpu

plot_object_probability()

cli argument --mdebug to dispatch the visualization

detect.py

I am still working on removing all hardcoding and simplifying the code more.

[glenn-jocher]

@seanavery oh that's really cool. Can you show some examples?

Yes, a debug mode would be interesting. If we can keep all of the plots seperated in general.py, or perhaps bring all plotting functions into their own plots.py, so the inline impact is minimal (i.e. in the loss function), then I think this could work well.

[seanavery]

object heatmap

[henbucuoshanghai]

@seanavery only output the heatmap? where is the box of instances?