Unsupervised CNN for Single View Depth Estimation: Geometry to the Rescue

Paper link: https://arxiv.org/pdf/1603.04992v2.pdf

• single view depth prediction
• unsupervised framework : without requiring pre-training stage or annotated ground-truth depth
• analogous to an autoencoder

Approach

• Make use of pairs of images with a known camera motion, such as stereo pairs.
• CNN: learn the complex non-linear transformation which converts the image to a depth-map
  • loss: photometric difference between the input (source image and the inverse warped image) differentiable and highly correlated tieh prediction error
• Interpreted in the context of convolutional autoencoeders

Autoencoder loss

• notation: $i \in \{1 \cdots N\}$: Training instance
• notation: $\{I_i^i, I_2^i\}$: Rectified stereo pair
• notation: $f$: focal length of the two cameras in a single pre-calibrated stereo rig, which capture the image pairs
• notation: $B$: horizontal distance beteen the cameras
• notation: $d^i(x)$: the predicted depth of a pixel $x$ in the left of the rig
• notation: $D^i(x) = fB/d^i(x)$ the motion of the pixel along the scan-line
• notation: $I_w^i = I^i_2()$

Inline math: $\int_{-\infty}^\infty g(x) dx$

Block math:

$\int_{-\infty}^\infty g(x) dx$