So, the idea is that I have a custom set of UAV images. I zoomed them and I used LabelImg in order to draw the rectangles. I trained the model and when I run the code for object detection it does detect the object it was trained for. My concern is that I used zoomed UAV images in order to train and unzoomed (original size images) images for object detection. Can this be the problem? What could be wrong? I used this code for train: https://tensorflow-object-detection-api-tutorial.readthedocs.io/en/latest/training.html and this code https://tensorflow-object-detection-api-tutorial.readthedocs.io/en/latest/auto_examples/plot_object_detection_saved_model.html#sphx-glr-auto-examples-plot-object-detection-saved-model-py for object detection.
4
-
1$\begingroup$ Did you mean that the model does NOT work as expected after training? Or what problem are you referring to? In case yes, the spatial resolution, i.e. meters per pixel, should be the same in training and inference. Not only that, if you train on images gathered from one place and test on images gathered on a different place that might also inficiate a lot the model performances. $\endgroup$– Edoardo GuerrieroAug 13, 2022 at 13:38
-
$\begingroup$ I don't see any rectangles in object detection/inference stage, on the outputted image... $\endgroup$– just_learningAug 13, 2022 at 13:44
-
1$\begingroup$ The yes, if you're sure about the code you got your culprit. Zoom also your images in inference phase, or even better, skip unnecessary preprocessing steps (resizing, no metter what type, always introduce artefacts), and train and infer on the original spatial resolution. $\endgroup$– Edoardo GuerrieroAug 13, 2022 at 13:50
-
$\begingroup$ How many image in your dataset? Do you use only drone image or other image (how many different label in your dataset)? $\endgroup$– Cloud ChoJan 11, 2023 at 20:34
Add a comment
|