Image Guided Depth Upsampling Using Anisotropic Total Generalized Variation
Summary (2 min read)
1. Introduction
- Accurate, high resolution depth sensing is a fundamental challenge in computer vision.
- Upsampling of a low resolution depth image (a) using an additional high resolution intensity image (b) through image guided anisotropic Total Generalized Variation (c). packet size and a low energy consumption make them applicable in mobile devices.
3.1. Depth Mapping
- Since the low resolution depth map DL and the high resolution intensity image IH stem from different cameras, a mapping can only be established when intrinsic and extrinsic parameters are known (see Section 4.2).
- In their setup the authors define the intensity camera as the world coordinate center.
- Each depth measurement di,j at pixel position xi,j = [i, j, 1] T is projected into the high resolution intensity image space ΩH.
- Therewith, the authors minimize the error which can occur due to this averaging in the high resolution space.
- Through the regularization term, introduced in Section 3.2, the area between the projected depth pixels is implicitly interpolated.
3.2. Depth Image Upsampling
- The authors upsampling method increases the resolution of measured depth data from a low resolution depth sensor by adding edge cues from a high resolution intensity image.
- (2) This formulation is composed of the data term G(u,DS) that measures the fidelity of the argument u to the input depth measurements DS and the regularization term F (u) that reflects prior knowledge of the smoothness of their solution.
- F and G are convex lower semi-continuous functions.
- Because the TGV regularizer is convex it allows to compute a globally optimal solution.
- Including this term in their TGV model the authors can penalize high depth discontinuities at homogeneous regions and allow sharp depth edges at corresponding texture differences.
3.3. Primal-Dual Optimization
- The proposed optimization problem (6) is convex but non smooth due to TGV regularization term and the zeros in the weighting operator w.
- To find a fast, global optimal solution for their problem the authors use the primal-dual energy minimization scheme, as proposed in [2, 6].
- The authors reformulate the non-smooth problem in a convex-concave saddlepoint problem applying the Legendre Fenchel transform (LF) .
- Second, the primal variables are updated using gradientdescent.
4. Evaluation
- The authors show a quantitative and qualitative evaluation of their upsampling method.
- (f) The authors upsampling method using image guided anisotropic TGV.
- (e) removes noise but suffers from edge bleeding especially at small structure boundaries.
4.1. Middlebury Benchmark Evaluation
- An exhaustive evaluation of their method in terms of quantitative and qualitative comparison is made using input images from the Middlebury datasets [10, 20].
- The authors use the disparity image as groundtruth and the original RGB intensity image as input for their anisotropic diffusion tensor.
- This experiment gives an objective comparison on the robustness, accuracy and speed of a variety of different algorithms.
- While the Middlebury datasets are popular to evaluate depth upsampling methods, they neglect some important properties of real acquisition setups.
- Typically, depth and intensity data do not originate from the same sensor and are therefore not aligned.
4.2. Benchmarking based on Real Sensor Data
- The evaluation on real acquisitions is made using different scenes acquired with a Time of Flight (ToF) and an intensity camera simultaneously.
- The rotation and translation between intensity and ToF camera is estimated by establishing a geometric correspondence through the feature points on the planar target.
- Through a comparison of the very accurate 3D measurements of the calibration points and the measured ToF depth points a dependence between the acquired IR amplitude image and the measurement error can be established, as shown in Figure 4.
- Using their depth calibration, the authors can compensate for that error (see green/dashed box).
- In the visual and numerical results it can be seen that their method delivers high quality upsampling results at multiple frames per second for an approximate upsampling factor of ×6.25.
5. Conclusion
- Low cost 3D sensor and an additional high resolution 2D sensor.the authors.
- The upsampling is formulated as a global energy optimization problem using Total Generalized Variation (TGV) regularization.
- For fast numerical optimization the authors use a first order primal-dual algorithm, which is efficiently parallelized resulting in high frame rates.
- In a quantitative evaluation using widespread datasets the authors show that their method clearly outperforms existing state of the art methods in terms of speed and quality.
- The authors further provide benchmarking datasets of real world scenes providing a highly accurate groundtruth that, for the first time, enable a real quality comparison of depth image upsampling methods.
Did you find this useful? Give us your feedback
Citations
518 citations
Cites background or methods or result from "Image Guided Depth Upsampling Using..."
...[12] perform slightly better than our method on very sparse data but require a dense high-resolution RGB image for guidance....
[...]
...Note that in contrast to other techniques [12, 52] which artificially upsample the input (e....
[...]
...We compare our unguided approach to several baselines [1, 12, 30, 58] which leverage RGB guidance for upsampling and two standard convolutional neural networks with and without valid mask concatenated to the input....
[...]
...More advanced approaches are based on global energy minimization [1, 6, 12, 49, 51], compressive sensing [22], or incorporate semantics for improved performance [58]....
[...]
353 citations
336 citations
Cites background or methods from "Image Guided Depth Upsampling Using..."
...Optimization algorithms of this nature have been used in global stereo [30], semantic segmentation [7, 20, 25, 38], depth superresolution [8, 17, 22, 24, 26, 27], and colorization [23]....
[...]
...With the advent of consumer depth sensors, techniques have been proposed for the task of upsampling the noisy depth maps produced by these sensors using a highresolution RGB reference image [8, 17, 22, 24, 26, 27]....
[...]
...The runtimes we report in Table 3 were either produced by ourselves (on a 2012 HP Z420 workstation) or taken from past work [8, 22]....
[...]
...To evaluate our model, we use a depth superresolution benchmark [8] which is based on the Middlebury stereo dataset [30]....
[...]
...Table 3: Performance on the depth superresolution task [8]....
[...]
317 citations
Cites methods from "Image Guided Depth Upsampling Using..."
...Figure S2 shows the convergence curves using f2 ∈ (3, 9, 11)....
[...]
300 citations
Cites methods from "Image Guided Depth Upsampling Using..."
...In applications using depth cameras, these complex systematic errors are usually calibrated and compensated as a preprocessing step before subsequence processing [44], [45]....
[...]
References
15,225 citations
Additional excerpts
...…according to⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨ ⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩ pn+1 = Pp { pn + σpα1 ( T 1/2(∇ūn − v̄n))} qn+1 = Pq {qn + σqα0∇v̄n} un+1 = un + τu ( α1∇TT 1/2pn+1 + wDS ) 1 + τuw vn+1 = vn + τv ( α0∇Tqn+1 + α1T 1/2pn+1 ) ūn+1 = un+1 + θ(un+1 − ūn) v̄n+1 = vn+1 + θ(vn+1 − v̄n) (10) until a stopping criterion is reached....
[...]
13,200 citations
"Image Guided Depth Upsampling Using..." refers background in this paper
...As a future perspective, it will be extended to incorporate a temporal coherence in a consistent way, eventually leading to depth reconstructions with even higher accuracy....
[...]
4,730 citations
"Image Guided Depth Upsampling Using..." refers methods in this paper
...The proposed method is not limited to single image upsampling....
[...]
4,487 citations
"Image Guided Depth Upsampling Using..." refers background or methods in this paper
...1550-5499/13 $31.00 © 2013 IEEE DOI 10.1109/ICCV.2013.127 993 We formulate the upsampling as a convex optimization problem [2, 6]....
[...]
...The gradient and divergence operators are approximated using forward/backward differences with Neumann and Dirichlet boundary conditions, respectively....
[...]
4,184 citations
"Image Guided Depth Upsampling Using..." refers methods in this paper
...To create whole volumes of depth data Newcombe et al ....
[...]