This readme explain how to setup the code-base used for trainining MAVREC benchmarks.
Please refer to Deformable DETR's github page for dependencies and other installations. We do not rely on anything additional.
We maintain distinct directories for the two views: Ground and Aerial samples. Our approach involves extracting samples from individual annotation files, which are then utilized to generate a combined annotation file encompassing both views. Instead of adding the directory path as a prefix to each image beforehand, we dynamically include it during training. We can do that because the image's file names describes the view. This requires a few steps seen below. Everything else is, as in the orignal code base.
- Generate a COCO annotation file containing both aerial and ground samples. File paths should only contain the base name.
- In following file:
datasets/torchvision_datasets/coco.pychange the__get_item__function's variables in the indentation belowif self.imageset == 'train'.
Thats it. Run everything else as normal.