- Provides a special insight into the working mechanisms of the human visual systems and brain.
- Aiming for neural network models to replicate visual illusions better can potentially improve their performance (Alexander 2021)
- Parallel lines are superimposed by oblique lines.
- The oblique lines are not parallel, but appear to be due to the effect of the parallel lines.
- To find the effect of Zöllner illusion on DNNs.
- Compare with human visual system.
- Predict the direction of the lines at each pixel.
- Direction is give by vector
$(x, y)$ - Input shape:
$3\times W\times H$ - Output shape:
$2\times W\times H$
- 3 CNNs
- 3 CNNs
-
$tanh$ activation layer
- 10 random lines without anti-aliasing
- Thickness ranges from 1 to 3 pixels
- Shape:
$3\times W\times H$
-
$(x,y)$ for each pixel, range from$[-1, 1]$ - Shape:
$2\times W\times H$
- line thickness ranges from 1 to 3 pixels
- slash direction ranges from 0 to 80 degrees
- Loss criterion
- Mean square error (MSE)
| Model | Test loss |
|---|---|
| Unet 64 | 427 |
| Unet 32 | 441 |
| Unet 16 | 635 |
| Unet 8 | 840 |
| VGG16 | 893 |
| Transformer | 525 |
From the result, we are able to tell that some DNNs is fooled by Zollner illusions just like human.
- Unet 64
- Perceive line segment tilting not uniformly but between slashes
- Sample outputs in HSV color representations
- Aqua (90 degrees in hue value); Blue (over 90 degrees on the left line)
- The results are consistent with the previous theory by Chiang
- Accurding to Chiang
- Due to light diffraction
- As two objects are closer, they merge into one maximum intensity
- Human visual systems
- Influence is strongest when the slash angle is 60 to 70 degrees
- DNNs
- U-net 64 and 32
- Inflence get stronger as the slash angle keep increasing
- Transformer
- Does not show consistency with human perception, even with high accuracy
- U-net 64 and 32