Running inference without datasets

[jama1017]

Hi,

Thanks for putting out this amazing work. I wonder if it is possible to directly use the pretrained model to run inferences on my own videos without going through the GUI or use eval.py, which requires me to download all the datasets? I guess my question is more about how I can do that if possible.

[hkchengrex]

eval.py does not require users to download all any dataset. You can structure your data as https://github.com/hkchengrex/XMem/blob/main/docs/INFERENCE.md#on-custom-data

And run python eval.py --output [your_output_directory] --dataset G --generic_path [your_data_root] --mem_every 10 (or with mem_every 5)

[jama1017]

Thanks so much for the clarification!

[jama1017]

Just as a follow up question - for the custom dataset, do we need to add annotation images for each of the video frames? Can we add no annotations at all? I tried the command you gave and it only processed the frames for which the annotations are provided.

[hkchengrex]

It should process all the frames since save_all should be true when dataset is G. I guess you can try passing --save_all but what you said is not expected. One annotation per video should be enough.

[jama1017]

Thanks for the quick reply. I figured out why - I didn't put in the annotation image for the first frame of the video but put in the annotation image for the last frame instead, so it only processed that last frame.