RF-DETR does not automatically resize input images during training (resolution parameter seems ineffective)

[universeliang]

Search before asking

• I have searched the RF-DETR issues and found no similar bug report.

Bug

I am encountering an issue when training RF-DETR with images of varying resolutions. My dataset contains images with different sizes, such as 960×540 and 1920×1080. Models like YOLO and RT-DETR can automatically resize images (e.g., to 640×640) during training by setting an image_size or similar parameter. However, in RF-DETR, I observed that the input images are not automatically resized, which leads to the following error:

AssertionError: Backbone requires input shape to be divisible by 32, but got torch.Size([8, 3, 1080, 1920])

I attempted the following approaches:

model.train(resolution=640) 

model = RFDETRNano(resolution=640)

However, neither of these had any effect. The input images were still passed to the model in their original resolutions, and the same error persisted.
Does RF-DETR support automatic image resizing during training (similar to YOLO or RT-DETR)?

Environment

• RF-DETR: 0.1.0
• OS: Windows 10
• Python: 3.10.0
• PyTorch: 2.5.1
• GPU: NVIDIA RTX 4070ti super

Minimal Reproducible Example

from rfdetr import RFDETRBase, RFDETRNano, RFDETRSmall, RFDETRMedium, RFDETRLarge

if __name__ == '__main__':
    model = RFDETRNano(resolution=640, num_classes=10)
    model.train(
        resolution=640
    )

Additional

No response

Are you willing to submit a PR?

• Yes, I'd like to help by submitting a PR!

[zhaoshuo1223]

Yes, it is supported. resolution=640is also not a problem. Initially, the images will be resized to 640. However, since your dataset includes images of 960 * 540, when I use it, the resolutionis typically set to be smaller than the minimum value in the dataset.

[JuMa555]

It is fixed by #956 and #964

[universeliang]

Yes, it is supported. resolution=640is also not a problem. Initially, the images will be resized to 640. However, since your dataset includes images of 960 * 540, when I use it, the resolutionis typically set to be smaller than the minimum value in the dataset.

I’m still a bit confused. I tried specifying the resolution parameter, but it doesn’t seem to solve the problem.

My dataset is VisDrone, which contains images with various resolutions. In models like YOLO or other DETR-based methods, the images are automatically resized to a fixed size (e.g., 640×640) at the beginning of training before being fed into the model. This effectively avoids issues caused by inconsistent input resolutions.

However, RF-DETR does not seem to provide such functionality. As a result, some images in my dataset (e.g., 1920×1080) cannot be processed because their dimensions are not divisible by 32, which causes the training to fail.

Could you clarify whether RF-DETR is supposed to handle resizing internally, or if users are expected to preprocess all images (e.g., resize or pad them) before training?

[lwdevelop191]

One solution is using resize as an augmentation.
"Resize": {"height": TRAIN_RES, "width": TRAIN_RES, "p": 1.0},

[Alarmod]

Why RF-DETR: 0.1.0? Use latest developer version

[universeliang]

Why RF-DETR: 0.1.0? Use latest developer version

I’m using the latest version; I just forgot to update the template when submitting the issue.

[zhaoshuo1223]

Yes, it is supported. resolution=640is also not a problem. Initially, the images will be resized to 640. However, since your dataset includes images of 960 * 540, when I use it, the resolutionis typically set to be smaller than the minimum value in the dataset.

I’m still a bit confused. I tried specifying the resolution parameter, but it doesn’t seem to solve the problem.

My dataset is VisDrone, which contains images with various resolutions. In models like YOLO or other DETR-based methods, the images are automatically resized to a fixed size (e.g., 640×640) at the beginning of training before being fed into the model. This effectively avoids issues caused by inconsistent input resolutions.

However, RF-DETR does not seem to provide such functionality. As a result, some images in my dataset (e.g., 1920×1080) cannot be processed because their dimensions are not divisible by 32, which causes the training to fail.

Could you clarify whether RF-DETR is supposed to handle resizing internally, or if users are expected to preprocess all images (e.g., resize or pad them) before training?

RF-DETR does not perform padding when processing data temporarily. When necessary, I usually do some preprocessing myself, such as padding, to convert the data into a square shape.
For images with various resolutions, I also encountered some problems after trying. #983