Would it help if I had more "ground truth" masks? about aot-benchmark HOT 2 OPEN

DeAOT/AOT supports using multiple reference frames during inference. The method is simple. You can place all the inference frames in the mask folder. During the inference, the algorithm will paste the reference frame to the predicted mask and cover your prediction.

I designed this function since, in some videos, some objects may not appear in the first frame but in one of the middle frames.

from aot-benchmark.

I designed this function since, in some videos, some objects may not appear in the first frame but in one of the middle frames.

@z-x-yang What about if the optimal reference frame is not frame 0? I see that the evaluator it is strictly forward only and it assume frame 0 reference frame.

from aot-benchmark.