Supplementary Material

Hi, there. If you are looking for demos, they are in section 5 and 6. Here are some shortcuts.
Each video is about 10MB. You may want to use WIFI instead of cellular data.

Plane      Car      Chair

Rifle      Table      Font

0. Paper, Code, Datasets and Pre-trained weights

Paper: https://arxiv.org/abs/1812.02822
Others: https://github.com/czq142857/implicit-decoder

1. Data preparation

See data-preparation.pdf

2. Network structures and training configurations

See network-structure.pdf

3. Toy experiments (extension of Figure 3)

See toy-experiments.pdf. The document provides several toy experiments to compare the different features learned by CNN-based models and implicit-decoder-based models.
Some animated interpolations (see more details at the document):
AE trained with CNN decoder:      AE trained with implicit decoder:
VAE trained with CNN decoder:      VAE trained with implicit decoder:
WGAN trained with CNN decoder:      WGAN trained with implicit decoder:

4. Auto-encoding 3D shapes (extension of Figure 4)

Each of the following images contains the visual comparisons of the first 16 shapes (sorted by name) from the testing set of that category.
plane      car      chair      rifle      table

5. 3D shape generation and interpolation (extension of Figure 6, videos of interpolations, etc.)

(a) 16 randomly generated shapes for each category for each model.
samples_3DGAN      samples_PC-GAN      samples_CNN-GAN      samples_IM-GAN256

(b) Samples comparing interpolations in IM-AE and IM-GAN latent spaces.
compare_AE_GAN_latent_space

(c) IM-GAN interpolation videos. Chair and table were trained at 64^3 resolution and others at 128^3. All shapes were retrieved using marching cubes at 256^3 resolution.
plane      car      chair      rifle      table

(d) CNN-GAN interpolation videos of chair and table for comparison (trained and sampled at 64^3).
chair      table

6. 2D shape generation and interpolation (extension of Figure 7 and videos of font interpolations)

(a) 1024 randomly generated samples for each model.
Not binarized:      DCGAN      CNN-GAN      IM-GAN      VAE_CNN      VAE_IM      WGAN_CNN      WGAN_IM
Binarized:            DCGAN      CNN-GAN      IM-GAN      VAE_CNN      VAE_IM      WGAN_CNN      WGAN_IM

(b) A video showing font interpolations with IM-GAN trained on 64^2 data and sampled at 128^2.
font

7. Single-view 3D reconstruction (extension of Figure 8 and evaluation results by chamfer distance)

(a) Each of the following images contains the visual comparisons of the first 16 shapes (sorted by name) from the testing set of that category.
plane      car      chair      rifle      table

(b) Evaluation results by chamfer distance
results_CD