I noticed that the camera poses in mvdiffusion/data/fixed_poses/nine_views (not exactly 0°,45°,90°,180°,270°,315°) is different from camera poses in instant-nsr-pl/datasets/fixed_poses (exactly 0°,45°,90°,180°,270°,315°). And the pictures inferenced from test_mvdiffusion_seq.py are a little inclined.
I also noticed the "self.camera_embedding" in mvdiffusion/pipelines/pipeline_mvdiffusion_image.py line 136 corresbonds to the views in mvdiffusion. It's also inclined.(for example, self.camera_embedding[2,2] should be half of pi(1.57) instead of 1.6934)
So how do the views in mvdiffusion folder come about?
I've changed the pose to multiples of 45°, and I get the correct results when training stage1 in the overfitting of a single object (6 views normals and rgbs).
Hi, Thanks for the wonderful job!
I noticed that the camera poses in mvdiffusion/data/fixed_poses/nine_views (not exactly 0°,45°,90°,180°,270°,315°) is different from camera poses in instant-nsr-pl/datasets/fixed_poses (exactly 0°,45°,90°,180°,270°,315°). And the pictures inferenced from test_mvdiffusion_seq.py are a little inclined.
I also noticed the "self.camera_embedding" in mvdiffusion/pipelines/pipeline_mvdiffusion_image.py line 136 corresbonds to the views in mvdiffusion. It's also inclined.(for example, self.camera_embedding[2,2] should be half of pi(1.57) instead of 1.6934)
So how do the views in mvdiffusion folder come about?