Use cotracker_stride_8_wind_16.pth And 6 frames get bad result

[Aatroy]

model: cotracker_stride_8_wind_16.pth
I modify the function named read_video_from_path in demo.py , just append 6 frames to video.
Then , I run the demo.py and print(pred_tracks), get the result below:
`

tensor([[[[465., 180.]],

     [[  0.,   0.]],

     [[  0.,   0.]],

     [[  0.,   0.]],

     [[  0.,   0.]],

     [[  0.,   0.]]]], device='cuda:0')

`
It seems no track result.

But when I append 10 frames to video , the result it's ok .
`
video = read_video_from_path(args.video_path)
points = torch.tensor([[0.,465.,180.]])

if torch.cuda.is_available():
    points = points.cuda()
video = torch.from_numpy(video).permute(0, 3, 1, 2)[None].float()


model = CoTrackerPredictor(checkpoint=args.checkpoint)
model = model.to(DEFAULT_DEVICE)
video = video.to(DEFAULT_DEVICE)

pred_tracks, pred_visibility = model(
    video,
    queries=points[None]
)
print(pred_tracks)
print("computed")`

[nikitakaraevv]

Hi @Aatroy, thank you for the question! This issue was discussed here: #40