How to find parameters when using my own camera images for inference，kitti360

[xym0601]

Thanks for your great job!
I successfully ran monoscene with my own images, but the results were not satisfactory. Is it because the following parameters need to be modified?If it needs to be modified, any suggestions?

self.root = root
self.img_H = 376
self.img_W = 1408
self.project_scale = 2
self.output_scale = 1
self.voxel_size = 0.2
self.vox_origin = np.array([0, -25.6, -2])
self.scene_size = (51.2, 51.2, 6.4)
self.T_velo_2_cam = self.get_velo2cam()
self.cam_k = self.get_cam_k()

How do I get these two matrices?
First of all, for the camera intrinsic parameter cam.k, I see that you are using the corrected intrinsic parameter instead of the original intrinsic parameter.
Is the cam2velo matrix still needed if it is only for inference?

    def get_cam_k(self):
        cam_k = np.array(
            [
                552.554261,
                0.000000,
                682.049453,
                0.000000,
                0.000000,
                552.554261,
                238.769549,
                0.000000,
                0.000000,
                0.000000,
                1.000000,
                0.000000,
            ]
        ).reshape(3, 4)
        return cam_k[:3, :3]

    def get_velo2cam(self):
        cam2velo = np.array(
            [
                0.04307104361,
                -0.08829286498,
                0.995162929,
                0.8043914418,
                -0.999004371,
                0.007784614041,
                0.04392796942,
                0.2993489574,
                -0.01162548558,
                -0.9960641394,
                -0.08786966659,
                -0.1770225824,
            ]
        ).reshape(3, 4)
        cam2velo = np.concatenate(
            [cam2velo, np.array([0, 0, 0, 1]).reshape(1, 4)], axis=0
        )
        return np.linalg.inv(cam2velo)

Hope to receive your reply, thank you.

[anhquancao]

You need to know your camera parameters. You can try to infer it using colmap if use have multi-view images.

[xym0601]

你需要知道你的相机参数。如果你有多视角图像，可以尝试使用 colmap 进行推断。

I only have one camera, so do I just need to modify the camera internal parameters?

[anhquancao]

Maybe you still need cam2velo to change the coordinates as the model was trained with cam2velo which make it learns the occupancy in the lidar coordinates. You can use colmap with multiple views from the same camera. You can try Dust3r also