I am working on ReIdentificationNet model based on Pytorch using NVIDIA TAO containers, I have some doubts about the dataset that we are using to train the ReidentificationNet model. Can someone help me out by sharing some informations about the person images in the dataset.
We know that the image sizes are 128X64, but what is entering the model is 256X128
1}Like we know that the person images in the dataset are present as cropped bboxes, are the aspect ratio of each person in every image is same, or is it different.
2} Is there any padding involved inside the person images becuase we can see that the images are captured from different angles then how it is possible that in every image the person inside the image is covering the same space becuase we are alsways passing 256X128 image through the model right?
Kindly help me.
Yes，Every image is resized to input_width x input_height.
About 10 pixels of padding(from the setting in yaml) are typically applied during training (for augmentation).
There is data augmentation techniques like horizontal flip or random erasing is applied to improve robustness. Refer to tao_pytorch_backend/nvidia_tao_pytorch/cv/re_identification/dataloader/transforms.py at main · NVIDIA/tao_pytorch_backend · GitHub
There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks.