Question about temporal dimension reversal in "Format._format_img" for Gen1 4D ".npy" inputs

[aaddaamm12]

Hi SpikeYOLO authors,

First, thank you for releasing the SpikeYOLO code and the Gen1-specific implementation. I am currently fine-tuning the Gen1 checkpoint on my own Prophesee neuromorphic dataset and I have a question about the image formatting pipeline for temporal ".npy" inputs.

In "SpikeYOLO_for_Gen1/ultralytics/data/augment.py", inside "Format._format_img", the current code is:

if len(img.shape) == 4:
     img = np.ascontiguousarray(img.transpose(0, 3, 1, 2)[::-1]) # 改为Tchw模式
else:
     img = np.ascontiguousarray(img.transpose(2, 0, 1)[::-1]) # 改为chw模式

For a standard OpenCV image in "HWC" format, I understand that: img.transpose(2, 0, 1)[::-1] converts "HWC -> CHW" and reverses the channel order, typically "BGR -> RGB".
However, for a temporal ".npy" input with shape:

[T, H, W, C]

the operation:

img.transpose(0, 3, 1, 2)[::-1]

first converts the image to:

[T, C, H, W]

but then "[::-1]" seems to reverse the first axis, which is now the temporal axis "T", not the
channel axis "C".

So, for example, with "T=2", this changes the temporal order from:

frame_0, frame_1

to:

frame_1, frame_0

Is what I am describing true or I am wrong?

If I'm right, my question is: Is this reversal of the temporal dimension intentional for the Gen1 model/checkpoint, or is it an inherited BGR-to-RGB operation from the original YOLO image pipeline?

I noticed this because my dataset is stored as ".npy" files with shape:

[T, W, H, C]

which is then converted by the loader BaseDataset.load_image to:

[T, H, W, C]

and then in Format._format_img, it is converted to

[T, C, H, W]

and I annotate the last temporal frame. When visualizing "train_batch0.png" during training, the displayed frame seemed to correspond to the first original frame rather than the last one. This raised a question for me. First, is the frame displayed in train_batch0.png only a debug visualization choice made by plot_images, or does it reflect the exact temporal order of the tensor that is also fed to the model during training and validation? In other words, if train_batch0.png shows the labels on the first original frame instead of the last annotated frame, should I interpret this as only a visualization issue, or as evidence that the model is also receiving the temporal frames in that reversed order?

After removing "[::-1]" from if len(img.shape) == 4:;img = np.ascontiguousarray(img.transpose(0, 3, 1, 2)[::-1]) # 改为Tchw模式 the visualization matched the expected annotated frame.

However, when fine-tuning from the released Gen1 checkpoint, I am not sure whether the checkpoint itself was trained with the original temporal reversal behavior. So I would like to know which behavior should be considered correct for fine-tuning the official Gen1 checkpoint:

1. Keep the current code because the Gen1 checkpoint was trained with this temporal
   reversal.
2. Remove "[::-1]" for temporal ".npy" inputs because reversing "T" is unintended.

Could you please clarify the intended behavior?

Thank you very much for your help.