Object detection is not giving the correct result

[ShahriarRahman]

Hi All,

I have been facing a challenge while detecting building footprints using the ArcGIS Image Analyst extension. I have trained the model, and the loss metrics and ground truth are showing a decent result, but when I am trying to detect the object, I am not getting the building footprints.

##################################################

Here are the details of the trained model,

{
"Framework": "arcgis.learn.models._inferencing",
"InferenceFunction": "[Functions]System\\DeepLearning\\ArcGISLearn\\ArcGISObjectDetector.py",
"ModelConfiguration": "_DynamicSSD",
"ModelType": "ObjectDetection",
"ExtractBands": [
0,
1,
2
],
"backbone": "resnet34",
"Grids": [
2,
4,
21
],
"Zooms": [
1.0
],
"Ratios": [
[
1.0,
1.0
]
],
"SSDVersion": 2,
"Classes": [
{
"Value": 1,
"Name": "build",
"Color": [
25,
66,
253
]
}
],
"MinCellSize": {
"x": 0.29992251483648985,
"y": 0.29993464576999285,
"spatialReference": {
"wkid": 32754,
"latestWkid": 32754
}
},
"MaxCellSize": {
"x": 0.29992251483648985,
"y": 0.29993464576999285,
"spatialReference": {
"wkid": 32754,
"latestWkid": 32754
}
},
"SupportsVariableTileSize": false,
"ArcGISLearnVersion": "1.8.5",
"monitored_valid_loss": 30.842981338500977,
"ModelFile": "model.pth",
"ImageHeight": 224,
"ImageWidth": 224,
"ImageSpaceUsed": "MAP_SPACE",
"LearningRate": "slice('1.0000e-02', '1.0000e-01', None)",
"ModelName": "SingleShotDetector",
"backend": "pytorch",
"ModelParameters": {
"backbone": "resnet34",
"backend": "pytorch"
},
"average_precision_score": {
"build": 0.0833333358168602
},
"resize_to": null,
"IsMultispectral": false
}

##################################################

This is the output I am getting (I have tried with different padding, thresholds and overlapping combinations),

Great to have your suggestions.

Kind regards,

Shahriar

[TimG]

Couple of things to check...

-The image is the same type is the training type, by type I mean like 8-bit unsigned.

-the cell size is consistent with trained data

As, at a guess, it could be that the training imagines are not 'equivalent' to the detect (inference)

[ShahriarRahman]

Hi @TimG,

Thank you for your response and suggestion. I have tried with the same image for both sampling and training, and also tried to run the model onto the same image.

Cell size is consistent with trained data and is 30 cm. Have already tried the settings you have shared. Still no success. I am not sure, but I think may be an issue with the deep learning framework. I am using ArcGIS Pro 2.8, and didn't install the python packages manually, but have used the installer package from Github. Don't know why the outcomes are not correct. 😞

Kind regards,

Shahriar

[TimG]

More thoughts. 

- average_precision_score doesn't look that high, looks to be 8% (?)

- I would have thought monitored_valid_loss should be way less

How many epochs are you doing?