May 2020
tl;dr: Improvement of neural net architecture to improve self-superv0sed depth. It leverage velocity when possible to achieve scale awareness.
The paper inherits all previous loss functions in self-supervised monocular depth method, and can be seen as a bag-of-trick for self-supervised mono depth.
The importance of high resolution has been demonstrated through SuperDepth ICRA 2019 via super-resolution with channel2spatial.
Most previous method requires depth scaling at test time. When velocity is available for regularization (during training), the network outputs metric results at minor degradation of performance. However it does not leverage this info during test time, and may fail at unseen ego motion. See SC-Sfmlearner and TrianFLow for other scale awareness solution.
- Use spatial2channel + 3D conv packing for more effective downsampling.
- The 3D packing + unpacking operation can preserve spatial details. Normal maxpool + bilinear upsample cannot keep original resolution. The code is relatively straightforward.
- The paper also adds a velocity loss when it is available.
- DepthNet Regress inverse depth.
- The proposed PackNet also performs better at higher resolution and does not plateau as early as ResNet.
- MR (640x192) vs HR (1280 * 384). Note that the ground and sky are cropped out.
- Github code
- The authors mentioned that it is possible to replace the poseNet prediction with GT pose, but they did not see significant improvement. github issue