search

dotnet / machinelearning

ML.NET is an open source and cross-platform machine learning framework for .NET.
https://dot.net/ml
MIT License
8.92k stars 1.86k forks source link

Auto ML Model Builder Object detection - bad or no results #6990

Open hannespreishuber opened 4 months ago

hannespreishuber commented 4 months ago

Windows 11 Surface Pro 8

I have a selfmade training set of free hand drawn simple objects - black on white - circle, rect, triangle. Training with AutoML image classification works fast and detection results are very good.

I did a dataset vott image tagging. Created a single json file and ~150 images. Training with AutoML and Object Detection. Takes ~25minutes and results are unusable. MAP50_95: 0,1258 Modell: ObjectDetectionMulti

Expected results is a better/good detection rate. I have no idea whats going on with my dataset. And how optimize it. Side effect: Error Message from Visual Studio "no bbox detected" have to restart VS to rerun..

attached Json and sample image 916a019c0896a9c18f5b7b2540b8014b-asset.json dreieck0

michaelgsharp commented 4 months ago

@LittleLittleCloud can you take a look at this since its run through automl?

LittleLittleCloud commented 4 months ago

@hannespreishuber model builder doesn't actually run AutoML optimizing on deep learning scenarios like object detection, so maybe you need to adjust hyper-parameters manually. For example, adding more epoches.

In the meantime, could you share dataset with us as well? It's possible that the poor metric is also dataset-related.

hannespreishuber commented 4 months ago

I am quite sure there is something I dont know in the dataset- My point is that Auto- doenst guide me in this vott-json-export.zip

my goal is to write training material for students