Allwinner v831 too low confidence on image classification when running resnet18

[Nurali125]

While working on resnet18, I found some interesting issues.

1. I custom trained resnet18 using transfer learning techniques for 10 classes. Then tested on my laptop using webcam (webcam is the same as on maix-ii dock board) and it showed good results. If I point webcam at a person, it results "people: 99.9%". Shows max confidence up to 100%.

Then, I converted this model to awnn form and uploaded to allwinner v831. Here is interesting - percentage of confidence was only between 15 - 45% no matter on what object (no threshold set, but anyway starts at around 15%). I tried to change focus of camera, bring the objects closer or further from camera, even showed images, on which the model was trained - it showed only this small range of confidence. But, on laptop, it was showing up to 100% confidence. Exact model, but converted to awnn showed much lower confidence.

2. Also, I looked at first 5 most confidence scores, first 2-3 objects were showing exact same percentage of confidence (e.g. airplane: 23.45%, car 23.45%, buildings 23.45%, road 3%, pheasant 0.23%). This is so weird, it must not show exactly the same confidence on any of objects.

3. I tried the model that was provided by allwinner (I run python app.py on therminal) --- just in rare cases it reached up to 70% confidence, but frequently ranged between 10 - 50% (10% threshold).

I could not find any reason for such behavior of resnet18 on allwinner v831, but I am thinking now is - a) either there are some bugs on converter pytorch - awnn OR b) there are bugs on libraries OR c) hardware performance

P.S. I attached link to videos and codes to make my words clearer. https://drive.google.com/drive/folders/17eKPOSqrP54RyBei_gemIOxNq8uV38Uo?usp=sharing

P.S. P.S. if you open sample_output folder, you can see all images of detected "airplanes". But, if you open "airplane.txt" file, you will see first 5 confidence scores of each image. Images are named: class + time + ".jpg". Each row of "airplane.txt" file is
named: time + date + confidence scores. Each row has its corresponding image (match them by time).

[Nurali125]

It's turned out that when you convert ncnn files to awnn using sipeed's website, part of dataset must be included in folder 'image' together with ncnn files. I thought that 'image' file is not so important, but in fact it is.