How to train Jetson Inference image classification with annotations in another PC

Robotics & Edge Computing Jetson & Embedded Systems Jetson Nano
1

Hello im using @dusty_nv jetson inference to do some image classification on the nano, Training on the nano though proved to be slow, so i went ahead and do the training on another computer using this repo still made by @dusty_nv and just import the resulting model to the nano. I have no problem training on the computer using that repo with the traditional way ( Creating a folder with test,val,train folders inside)

But i would like to train images that has annotation for classification. Is that even possible for classification? As per this thread it seems to be possible for detection.

I use CVAT to annotate the images, for detection it seems to use the PASCAL format, but i am not sure what to use for imagenet (There is a imagenet format in CVAT). if possible i would still like to train on another computer and just import the resulting model to the jetson

3

Exporting to IMAGENET in CVAT has gave me this file structure

Folder
   -->Combined Images
           --> Class A Image1.jpg
           --> Class B Image1.jpg
           --> Class C Image1.jpg
           .  .  .
           --> Class Z Image1.jpg
   -->Class A
           --> Class A Image1.jpg
           .  .  .
           --> Class A Image10.jpg
   -->Class B
           --> Class B Image1.jpg
           .  .  .
           --> Class B Image10.jpg
   .  .  .
   -->Class Z
           --> Class Z Image1.jpg
           .  .  .
           --> Class Z Image10.jpg

Not even some kind of file that represents the annotations somehow? Is there a specific way on how i should annotate the images?

That was exporting with “save images” enabled. Exporting again and unchecking this options, it provided me with 2 files

synset.txt
default.txt

synset.txt seems to contain the labels classification labels, while default.txt seems to contain the filename of the images and some numbers.

Is this workable? Because it is still using a different structure to what i used for training imagenet.

This is the current working folder structure:

Folder
   ListOfLabels.txt
   -->Train
           --> Class A
                 -->Class A Image1.jpg
           --> Class B
                 -->Class B Image1.jpg
           .   .   .
           --> Class Z
                 -->Class B Image1.jpg
   -->Val
           --> Class A
                 -->Class A Image11.jpg
           --> Class B
                 -->Class B Image11.jpg
           .   .   .
           --> Class Z
                 -->Class B Image11.jpg
   -->Test
           --> Class A
                 -->Class A Image50.jpg
           --> Class B
                 -->Class B Image50.jpg
           .   .   .
           --> Class Z
                 -->Class B Image50.jpg
4

@thesis.finals333 there are no separate annotation files for classification, those are simply determined from the dataset’s directory structure. By annotation I think you mean object detection, which contain the bounding box info in separate files (i.e. in Pascal VOC format)

5

I see,

I saw the imagenet tagging thread i thought i was possible:

No worries though as i only have 1 object in my images

7

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.