This content will become publicly available on December 16, 2024
- Award ID(s):
- 2107077
- PAR ID:
- 10545648
- Publisher / Repository:
- Conference on Neural Information Processing Systems
- Date Published:
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
Unsupervised monocular depth estimation techniques have demonstrated encour- aging results but typically assume that the scene is static. These techniques suffer when trained on dynamical scenes, where apparent object motion can equally be ex- plained by hypothesizing the object’s independent motion, or by altering its depth. This ambiguity causes depth estimators to predict erroneous depth for moving objects. To resolve this issue, we introduce Dynamo-Depth, an unifying approach that disambiguates dynamical motion by jointly learning monocular depth, 3D independent flow field, and motion segmentation from unlabeled monocular videos. Specifically, we offer our key insight that a good initial estimation of motion seg- mentation is sufficient for jointly learning depth and independent motion despite the fundamental underlying ambiguity. Our proposed method achieves state-of-the-art performance on monocular depth estimation on Waymo Open [34] and nuScenes [3] Dataset with significant improvement in the depth of moving objects. Code and additional results are available at https://dynamo-depth.github.io.more » « less
-
Unsupervised monocular depth estimation techniques have demonstrated encouraging results but typically assume that the scene is static. These techniques suffer when trained on dynamical scenes, where apparent object motion can equally be explained by hypothesizing the object's independent motion, or by altering its depth. This ambiguity causes depth estimators to predict erroneous depth for moving objects. To resolve this issue, we introduce Dynamo-Depth, an unifying approach that disambiguates dynamical motion by jointly learning monocular depth, 3D independent flow field, and motion segmentation from unlabeled monocular videos. Specifically, we offer our key insight that a good initial estimation of motion segmentation is sufficient for jointly learning depth and independent motion despite the fundamental underlying ambiguity. Our proposed method achieves state-of-the-art performance on monocular depth estimation on Waymo Open and nuScenes Dataset with significant improvement in the depth of moving objects. Code and additional results are available at https://dynamo-depth.github.io.more » « less
-
more » « less
Abstract In the vegetation root zone, infiltration (Inf) parts in two directions with distinct Earth-system functions. One goes up as evapotranspiration (E + Tr), returning Inf to the atmosphere (short-circuiting) and affecting short-term weather/climate and the carbon cycle. The other goes down as deep drainage (DD), flushing the regolith, mobilizing nutrients/contaminates and dissolved minerals into aquifers and rivers, eventually reaching the ocean (long-circuiting) thus regulating global biogeochemical cycles and long-term climate. We ask, what is the modern-day global structure in short- vs. long-circuiting? What forces and feedbacks create such structures? Synthesizing site-studies aided by global modeling, we found that: (i) long-circuiting prevails in evenly wet climates, in well-drained landscapes with a deep vadose zone, in substrates with deep conduits, and with plant biomass below natural equilibrium; (ii) soil B-horizons, via geochemical and vegetation feedbacks, enhance short-circuiting, while deep rock fractures enable long-circuiting even in dry climates; (iii) in dry climate/season and in uplands, plant roots follow Inf into deep vadose zone to tap wet-season Inf; (iv) plant water-use reinforces shallow Inf, reducing DD and regolith flushing in dry and season-dry climates; (v) where short-circuiting prevails, a dry soil zone separates modern surface processes from fossil groundwater; and (vi) the E + Tr supply depth, regolith flushing rate, and groundwater residence time vary greatly across the land, arising from multiscale drivers/feedbacks among climate, drainage, substrate, and biomass. These findings link site-based process discoveries to Earth-system level structures and functions of water belowground, shedding light on where/when/how the infiltrated rain influences the atmosphere above or the ocean downstream.
-
more » « less
Single-photon lidar (SPL) is a promising technology for depth measurement at long range or from weak reflectors because of the sensitivity to extremely low light levels. However, constraints on the timing resolution of existing arrays of single-photon avalanche diode (SPAD) detectors limit the precision of resulting depth estimates. In this work, we describe an implementation of subtractively-dithered SPL that can recover high-resolution depth estimates despite the coarse resolution of the detector. Subtractively-dithered measurement is achieved by adding programmable delays into the photon timing circuitry that introduce relative time shifts between the illumination and detection that are shorter than the time bin duration. Careful modeling of the temporal instrument response function leads to an estimator that outperforms the sample mean and results in depth estimates with up to 13 times lower root mean-squared error than if dither were not used. The simple implementation and estimation suggest that globally dithered SPAD arrays could be used for high spatial- and temporal-resolution depth sensing.
