Deep Convolutional Compressed Sensing for LiDAR Depth Completion

• The task is to convert sparse LiDAR or stereo depth samples into high-resolution depth maps for robotic sensing.Source: §1 · p. 1
• Existing networks treat completion as sparse-to-dense mapping, often fail to propagate input points, and remain sensitive to input sparsity.Source: §1 · p. 1
• Compressed sensing provides a framework for recovering signals from incomplete measurements using sparse representations in an overcomplete basis.Source: §1 · p. 2
• Multi-layer convolutional sparse coding models local interactions efficiently, but existing methods were not designed to learn from sparse ground-truth depth data.Source: §1 · p. 2
• CNNs with ReLU activations are related to layered thresholding for convolutional sparse coding, motivating the paper’s ADNN-based architecture.Source: §1 · p. 2
• The proposed network encodes a more sophisticated ML-CSC algorithm and outperforms published KITTI depth-completion results with far fewer parameters and layers.Source: §1 · p. 3
• Using two dictionary layers and 1600 parameters, the method substantially outperforms deep networks with more than 20 layers and over 3 million parameters.Source: §1 · p. 3

2 Related Work

• The related-work section reviews compressed sensing and deep learning approaches because the proposed method combines both techniques.Source: §2 · p. 3

2.1 Compressed Sensing

• Earlier compressed-sensing methods used single-level handcrafted wavelet, contourlet, or scene-specific dictionaries for depth reconstruction.Source: §2.1 · pp. 3–4
• Unlike these methods, the paper learns multi-level convolutional dictionaries from large datasets of incomplete ground-truth depth maps.Source: §2.1 · p. 4

2.2 Deep Learning

• Deep-learning depth estimation largely began with RGB-to-depth prediction, while LiDAR completion work introduced sparse-input-aware CNNs and established the comparison baselines.Source: §2.2 · p. 4
• The paper uses the Sparse-to-Dense and Sparsity Invariant CNN methods as state-of-the-art LiDAR depth-completion baselines.Source: §2.2 · p. 4

3.1 Compressed sensing

• Compressed sensing reconstructs a full-resolution depth map from sparse LiDAR measurements using a sparse linear combination of an overcomplete dictionary.Source: §3.1 · p. 4
• The code and dictionary produce the predicted depth map, while a measurement mask restricts reconstruction error to observed pixels.Source: §3.1 · p. 4
• The formulation addresses how to design a dictionary for applying CS in Eq. 1 to high-dimensional signals such as depth maps.Source: §3.1 · p. 4
• Factoring the dictionary into multi-layer convolutional structure enables an unrolled recurrent network that learns dictionaries and hyperparameters end to end.Source: §3.1 · p. 5

3.2 Deep Component Analysis

• Multi-layer sparse coding requires sparse intermediate representations, making large effective dictionaries tractable while increasing receptive fields with manageable parameter counts.Source: §3.2 · p. 5
• ADNNs solve sparsity-regularized objectives through differentiable ADMM iterations, interpreting layer activations as proximal operators and learning parameters by backpropagation.Source: §3.2 · p. 5

4.1 Inference

• Masked convolutional sparse coding treats extracting higher-level codes from sparse depth as a missing-data problem, with masks propagated through convolutional receptive fields.Each mask records whether a code element has valid inputs, computed by max pooling; the objective combines depth and code reconstruction costs.Source: §4.1 · p. 6
• ADMM alternates updates of sparse codes, auxiliary variables, and dual variables, using incremental layerwise updates that are reported to converge faster.The method constrains zi = yi and minimizes the augmented Lagrangian over variables in turn.Source: §4.1 · p. 6
• A Parseval tight-frame assumption replaces the expensive dictionary-related matrix inversion with a computationally cheaper differentiable update.The inversion is costly because Wi is large; the approximation assumes WiWi^T = I and enables a binomial-identity rewrite.Source: §4.1 · p. 7
• The unrolled ADMM procedure is implemented with convolution, transpose-convolution, and ReLU layers, producing optimized high-level codes whose effective dictionary reconstructs the depth map.Unlike a standard decoder, reconstruction is taken from the final code because its elements have the largest receptive field and fewest missing entries.Source: §4.1 · pp. 7–8

4.2 Learning

• Training learns dictionaries and biases by stochastic gradient descent with masked reconstruction loss, while TV-L1 smoothness improves visual quality without significantly changing quantitative errors.The ground-truth mask excludes invalid pixels, and α = 0.1 was selected for the smoothness weighting.Source: §4.2 · pp. 8–9

5.1 Implementation Details

• The implementations use one, two, or three convolutional sparse-coding layers, with ADMM iterations fixed or varied according to the experiment.The two- and three-layer variants use ten training and testing iterations except in Section 5.4.Source: §5.1 · p. 9

Error Metrics

• The evaluation uses RMSE, MAE, MRE, and tighter inlier thresholds δ_i with thresholds 1.01^i rather than 1.5^i.The benchmark uses accumulated and filtered LiDAR measurements to provide quality ground truth for dense-depth recovery from a single sweep.Source: Error Metrics · p. 9, §5.2 · p. 9

5.2 KITTI Depth Completion Benchmark

• Our method outperforms all previous state-of-the-art depth-only and RGB-guided methods on the KITTI Depth Completion benchmark.Table 1 reports validation errors for the compared methods.Source: Table 1 · p. 10, §5.2 · p. 10

Varying Sparsity Levels

• With only 50% of input LiDAR samples retained, the method still outperforms both baselines, although SparseConvNet has greater sparsity invariance.The keep probability denotes the probability that each LiDAR sample is retained.Source: Varying Sparsity Levels · p. 10, Figure 2 · p. 11

5.3 Effect of Amount of Training Data

• The method outperforms both baselines for every training-set size, matches the sparsity-invariant network with 100 examples, and improves with more data.Ma & Karaman’s 3.4M-parameter model generalizes poorly on small datasets, while Uhrig et al.’s method gains no performance from additional data.Source: §5.3 · p. 11, Figure 3 · p. 11

5.4 Effect of Iterative Optimization

• Iterative ADMM refinement is central: few iterations fail to enforce constraints, while increasing iterations produces a substantial performance boost.The network seeks depth maps consistent with sparse-input constraints and possessing sparse representations.Source: §5.4 · p. 12, Figure 4 · p. 12

6 Conclusion

• The proposed deep recurrent autoencoder establishes a link between sparse coding and deep learning for depth completion.The architecture builds on Murdock et al.’s Deep Component Analysis.Source: §6 · p. 12
• The model outperforms existing depth-completion methods, including methods that leverage RGB information.The conclusion also attributes success to internal optimization and the model’s small number of parameters.Source: §6 · pp. 12–13
• Selected KITTI visual results are arranged as RGB image, ground truth, input LiDAR points, and predicted depth from top to bottom.The figure provides qualitative examples rather than a numerical comparison.Source: Figure 5 · p. 14