Run with calibrated videos

[Jianhao-zheng]

Hi, thank you for sharing this amazing work!

I want to run Vipe on some SLAM benchmark dataset, which has calibrated intrinsic information. I made several small modifications to make it run with known intrinsics. I would appreciate if you can help confirm my implementation is sufficient.

1. I added the following parameters to configs/slam/default.yaml

cam: 
  fx: ...
  fy: ...
  cx: ...
  cy: ...

2. In this line:

   vipe/vipe/pipeline/default.py

   Line 67 in 71f2cc7

   init_processors.append(GeoCalibIntrinsicsProcessor(video_stream, camera_type=self.camera_type))

   the added camera intrinsic parameters (self.slam_cfg.cam) are then parsed as an additional argument to GeoCalibIntrinsicsProcessor, and the parameter dict will be set as its attribute self.cam_cfg.
3. I further modify the __call__ function here

   vipe/vipe/pipeline/processors.py

   Line 56 in 71f2cc7

   def __call__(self, frame_idx: int, frame: VideoFrame) -> VideoFrame:

   
   to:

def __call__(self, frame_idx: int, frame: VideoFrame) -> VideoFrame:
        frame.intrinsics = torch.as_tensor(
            [self.cam_cfg.fx, self.cam_cfg.fy, self.cam_cfg.cx, self.cam_cfg.cy]
        ).float()
        frame.camera_type = self.camera_type
        return frame

4. Set the parameter pipeline.slam.optimize_intrinsics as false.

This approach seems to work, as the output intrinsics match my config. Is this the intended and sufficient way to use fixed intrinsics, or is there a better method?

Thanks for your help!