Learning for 3D Vision - HW1¶

Name: Arnav Samal, Andrew ID: asamal

1¶

1.1¶

GIF of cow mesh from many continuously changing viewpoints

1.2¶

GIF of Cow with Dolly Zoom Effect

2¶

2.1¶

360-degree GIF animation of tetrahedron

There are 4 vertices and 4 triangle faces

2.2¶

360-degree GIF animation of cube

There are 8 vertices and 2 x 6 = 12 triangle faces

3¶

360-degree GIF animation of cow but with my colour gradient

Chose the Golden State Warrior's colors (blue and yellow):

  • Blue: [0.00, 0.29, 0.58]
  • Yellow: [1.00, 0.78, 0.00]

4¶

R_relative and T_relative follow PyTorch3D's camera extrinsic convention (row-vector convention)

i.¶

First Transform of Cow Mesh

  • R_relative = torch.tensor([ [ 0., 1., 0.], [-1., 0., 0.], [ 0., 0., 1.] ])

  • T_relative = torch.tensor([0., 0., 0.])

R_relative rotates the camera 90 degrees about the z-axis, while T_relative applies no additional translation. Final image looks like the camera is flipped

ii.¶

Second Transform of Cow Mesh

  • R_relative: tensor([ [1., 0., 0.], [0., 1., 0.], [0., 0., 1.]])
  • T_relative: tensor([0., 0., 2.])

R_relative is an identity matrix here so it performs no rotation, while T_relative applies a translation of 2 along the z-axis (towards negative), pushing the camera farther away. Final image looks zoomed out from the original

iii.¶

Third Transform of Cow Mesh

  • R_relative: tensor([ [1., 0., 0.], [0., 1., 0.], [0., 0., 1.]])
  • T_relative: tensor([ 0.5000, -0.5000, 0.0000])

R_relative is an identity matrix here so it performs no rotation, while T_relative translates the camera by -0.5 in x-axis and +0.5 in y-axis, shifting the object slightly left and below the image center.

iv.¶

Fourth Transform of Cow Mesh

  • R_relative: tensor([ [ 0., 0., 1.], [ 0., 1., 0.], [-1., 0., 0.]])
  • T_relative: tensor([-3., 0., 3.])

R_relative rotates the camera 90 degrees about the y-axis and T_relative helps moves it to the negative x-axis (with T_0 being [0, 0, 3]).

5¶

5.1¶

No description has been provided for this image

First Point Cloud

No description has been provided for this image

Second Point Cloud

No description has been provided for this image

Union of First and Second Point Clouds

(Sampled every 4 points due to rendering constraints)

5.2¶

360 degree animation of Torus Shape

360 degree animation of Torus Shape

360 degree animation of Waves

360 degree animation of Waves

(Sampled every 4 points due to rendering constraints for both shapes)

5.3¶

360 degree animation of Torus mesh

360 degree animation of Torus Mesh

  • Rendering speed: Point clouds are usually faster to render because they only have points, while meshes have faces that need to be processed.
  • Rendering quality: Meshes usually look better and smoother because they represent the full surface. Point clouds can look more scattered, especially with fewer points.
  • Ease of use: Point clouds are simpler because we only need a set of 3D points. Meshes are a little more complicated because they also need faces.
  • Memory usage: Point clouds can use less memory when there are fewer points, but using a lot of points can increase memory usage. Meshes also use memory for both vertices and faces.
  • Overall: Meshes are better when we want a complete and smooth 3D object, while point clouds are better when we want something simple and easy to sample or work with.

360 degree animation of Diamond mesh

360 degree animation of Diamond Mesh

6¶

I really like airplanes (a bit of an avgeek) so I decided to include an airplane here. Source: Free3D Link

360 degree animation of Airplane

360 degree animation of (looks like) Cessna 150

Had to work with the meshes a bit (rotate and normalize them), especially the textures, to get everything to fit properly. Rendering took around 2 hours, so I kept the quality low.

7¶

360 degree animations of the cow mesh and sampled point clouds

From left to right: original cow mesh, 10, 100, 1000, and 10000 sampled point clouds.