Incorrect Alignment When Rendering Views with Custom FOV and Camera Radius

Hi, thanks for the great work on SAM3D-Object!
I encountered an issue when trying to reproduce multi-view consistency using my own Blender-rendered images.

🐛 Issue Description

I rendered an object in Blender with:

FOV = 49.1

Camera radius = 1.5

Views: azimuths = 0°, 90°, 180°, 270°, elevation = 0°

Then:

I used the azimuth = 0° image as input to sam3d-object to generate the 3D reconstruction.

I attempted to render the generated object from azimuths 90°, 180°, 270° using the notebook’s default settings:

FOV = 60

r = 1

→ The rendered views do not align with my original Blender images.

I also tried rendering with my original parameters (FOV = 49.1, r = 1.5) inside the SAM3D-Object rendering pipeline.

→ The output still fails to align with the original rendered images.

The rendered results is too big compare to the gt input and render images

🔍 Rendering Code Used
# yaw = azimuth
yaws = [math.radians(a + azi_bias) for a in azimuth_list]

# pitch = elevation
pitchs = [math.radians(elevation_deg)] * len(azimuth_list)

extr, intr = _yaw_pitch_r_fov_to_extrinsics_intrinsics(yaws, pitchs, r, fov)

out = render_utils.render_frames(
sample,
extr,
intr,
{"resolution": resolution, "bg_color": bg_color, "backend": "gsplat"},
**kwargs,
)
return out # out["color"] shape = [N, H, W, 3]

And the inference workflow:

from utils.inference import Inference, render_video, make_scene

outputs = [self.inference(image_rgb, mask, seed=seed) for mask in masks]

# 5. Merge scene and prepare for rendering
scene_gs = make_scene(*outputs)
scene_gs_ready = ready_gaussian_for_video_rendering(scene_gs)

❓ Question

Is there a known mismatch between Blender's camera model and the _yaw_pitch_r_fov_to_extrinsics_intrinsics conversion used in SAM3D-Object?

Or is there an additional transformation needed to ensure the rendered views match Blender’s coordinate conventions?

Any guidance on how to correctly align the camera parameters would be greatly appreciated!