Depth on Demand: Streaming Dense Depth from a Low Frame Rate Active Sensor
-------
Interactive Supplementary Material

In the following sections we provide a set of video and interactive qualitative examples of our framework working in different scenarios. In each example the sparse depth points are largely dilated for visualization purposes.

ScannetV2 Dataset

In this section we provide examples of mesh reconstruction exploiting our framework, in each scene we assume the availability of a few active sensor frames, while the camera has an higher frame rate. On the left can be qualitatively evaluated the online 3D reconstruction of the scene. On the right the 3D mesh reconstruction can be further investigated in the interactive mesh visualizer (left-click to rotate, right-click to translate). The visualized scene can be changed using the button below.

7Scenes Qualitatives

Generalization results are obtained on the 7Scenes dataset training on ScannetV2. Our approach is still effective on unknown envinronments.

TartanAir Qualitatives

Here we show a set of video samples from TartanAir. From left to right the source view, the target view with the projected sparse depth points, the predicted depth and the ground-truth. The sparse depth is updated every 5 frames. Sparse depth points are largely dilated for visualization purposes.

KITTI Qualitatives

Here we show a set of video samples from KITTI. From left to right the source view, the target view with the projected sparse depth points, the predicted depth and the ground-truth. The sparse depth is updated with a 1Hz frequency. It's worth mentioning that our framework smoothly deals with reprojection errors due to moving objects as can be observed in the examples. Sparse depth points are largely dilated for visualization purposes.