UnDEMoN 2.0: Improved Depth and Ego Motion Estimation through Deep Image Sampling

V. Madhu Babu, Swagat Kumar, Anima Majumder, Kaushik DasarXiv 2018 · OpenReview

Abstract

In this paper, we provide an improved version of the UnDEMoN model for depth and ego motion estimation from monocular images. The improvement is achieved by combining the standard bi-linear sampler with a deep network based image sampling model (DISNet) to provide better image reconstruction capabilities, on which depth estimation accuracy depends in unsupervised learning models. While DISNet provides higher-order regression and a larger input search space, the bi-linear sampler provides geometric constraints necessary for reducing the size of the solution space for this ill-posed problem. This combination is shown to provide significant improvement in depth and pose estimation accuracy, outperforming existing state-of-the-art methods. Additionally, the modified network uses far fewer tunable parameters, making it one of the lightest deep network models for depth estimation. The efficacy of UnDEMoN 2.0 is demonstrated through rigorous experimental analysis on the standard KITTI dataset.

Introduction

Estimating depth and ego-motion from monocular images is a fundamental challenge in autonomous driving and robotics. While supervised methods yield high accuracy, they rely on expensive ground-truth data like LiDAR. Unsupervised methods circumvent this by using view synthesis as a supervisory signal. However, these methods typically rely on bi-linear interpolation (BIS) to warp images during training. BIS only considers a small local neighborhood of pixels, which limits its effectiveness in handling large displacements, complex textures, and distant objects. This "sampling bottleneck" often leads to blurred reconstructions and degraded depth predictions. UnDEMoN 2.0 addresses this by introducing a hybrid sampling approach that augments traditional geometry with deep learning.

Methodology

The UnDEMoN 2.0 framework utilizes a tripartite architecture designed for efficiency and accuracy. It consists of three primary components:

  1. DispNet: An encoder-decoder network that predicts disparity maps from single monocular images.
  2. PoseNet: A network that estimates the 6-degree-of-freedom camera motion between consecutive frames.
  3. DISNet (Deep Image Sampler Network): A novel component that performs higher-order regression for image reconstruction, overcoming the limitations of simple interpolation.
The UnDEMoN 2.0 architecture featuring DispNet, PoseNet, and DISNet for unsupervised training.

The core innovation lies in the hybrid sampling strategy. While the standard bi-linear sampler (BIS) enforces rigid geometric constraints to keep the solution space manageable, DISNet provides a larger search space and higher-order interpolation. This allows the model to reconstruct structural details—such as sharp edges and fine textures—that mathematical interpolation often misses. The training process is guided by a composite loss function including appearance loss (SSIM and L1 norm with Charbonnier penalty), disparity smoothness, and left-right consistency.

Results and Discussion

The model was rigorously evaluated on the KITTI dataset using the Eigen and Stereo splits for depth, and the Odometry split for ego-motion.

Performance and Efficiency

UnDEMoN 2.0 demonstrates a significant leap in performance over its predecessor and other state-of-the-art models. It achieved a 4.47% improvement in depth accuracy over the original UnDEMoN and up to an 18% improvement over models like UnDeepVO. Remarkably, this performance gain comes with a drastic reduction in model size. By pruning unnecessary layers (specifically cnv6 and cnv7), the authors reduced the parameter count to approximately 8 million, which is only 25% of the original model's size and significantly lighter than competitors like Monodepth (~32M parameters).

Qualitative Analysis

The impact of DISNet is most visible in the quality of reconstructed images. While BIS tends to produce blurry results in high-frequency areas, DISNet excels at preserving structural attributes.

Qualitative comparison between Bilinear Image Sampling (BIS) and Deep Image Sampling (DIS).

As shown in the comparison, DISNet provides much clearer reconstructions of object boundaries and distant structures. This cleaner reconstruction signal translates directly into more accurate gradients for the DispNet, leading to superior depth maps. Furthermore, the improved depth estimation directly benefits ego-motion tracking, resulting in lower Absolute Trajectory Error (ATE) compared to SfMLearner.

Conclusion

UnDEMoN 2.0 represents a significant advancement in unsupervised monocular depth estimation. By integrating a deep image sampler (DISNet) with traditional geometric constraints, the model achieves state-of-the-art accuracy while remaining exceptionally lightweight. A current limitation is the static scene assumption, which can lead to errors in the presence of independently moving objects. Future research may explore the integration of optical flow or adversarial training to better handle dynamic environments and further enhance reconstruction quality.

Summary updated

Abstract

In this paper, we provide an improved version of UnDEMoN model for depth and ego motion estimation from monocular images. The improvement is achieved by combining the standard bi-linear sampler with a deep network based image sampling model (DIS-NET) to provide better image reconstruction capabilities on which the depth estimation accuracy depends in un-supervised learning models. While DIS-NET provides higher order regression and larger input search space, the bi-linear sampler provides geometric constraints necessary for reducing the size of the solution space for an ill-posed problem of this kind. This combination is shown to provide significant improvement in depth and pose estimation accuracy outperforming all existing state-of-the-art methods in this category. In addition, the modified network uses far less number of tunable parameters making it one of the lightest deep network model for depth estimation. The proposed model is labeled as "UnDEMoN 2.0" indicating an improvement over the existing UnDEMoN model. The efficacy of the proposed model is demonstrated through rigorous experimental analysis on the standard KITTI dataset.
BibTeX
@misc{babu2018undemon,
  author = {V. Madhu Babu and Swagat Kumar and Anima Majumder and Kaushik Das},
  title = {{UnDEMoN 2.0: Improved Depth and Ego Motion Estimation through Deep Image Sampling}},
  year = {2018},
  eprint = {1811.10884},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1811.10884},
}
Figure