how to use the deeplabv3 model? about flutter-tflite HOT 2 CLOSED

When you get 21 values back, what do they look like for something like pixel 0,0? If I remember correctly, there should only be 20 values in deeplabv3:

aeroplane
bicycle
bird
boat
bottle
bus
car
cat
chair
cow
diningtable
dog
horse
motorbike
person
pottedplant
sheep
sofa
train
tv

The way I've used it in the past for an Android project is to loop through the returned arrays and grabbing that last index to treat it as an index of a list of labels (manually adding the labels to an array rather than associating a labels metadata with it). I'm curious if there's maybe a confidence level that you're getting for that last value set or something that I'm not catching :)

from flutter-tflite.

@PaulTR Thanks a lot for your response. I have finally resolved the issue. The provided list of labels is accurate. However, upon reviewing the Android example, I noticed that the first label corresponds to the background. Therefore, we have a total of 21 labels, including the background label.

Based on my assumption, these values represent the scores. The highest score indicates that the pixel belongs to the respective label. In the given example, we can infer that it represents a person.

16.766469955444336 // 0 => background
6.716494560241699
5.2359299659729
7.035886764526367
9.835386276245117
14.302803039550781
7.088613510131836
6.11181640625
2.46304988861084
8.762212753295898
7.735017776489258
11.51220417022705
8.09216594696045
7.704705238342285
-1.0880954265594482
22.918304443359375 // 15 => person
4.772500038146973
2.029364824295044
1.225829839706421
4.1064910888671875
8.370333671569824

from flutter-tflite.