Unsupervised Learning of Monocular Depth and Ego-Motion using Conditional PatchGANs
Authors: Madhu Vankadari, Swagat Kumar, Anima Majumder, Kaushik Das
IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The efficacy of the proposed approach is demonstrated through rigorous ablation studies and exhaustive performance comparison on the popular KITTI outdoor driving dataset. |
| Researcher Affiliation | Industry | Madhu Vankadari , Swagat Kumar , Anima Majumder and Kaushik Das TCS Research and Innovation, Bangalore, India {madhu.vankadari, swakat.kumar, anima.majumder and kaushik.da}@tcs.com |
| Pseudocode | No | No pseudocode or algorithm blocks were found. The network architectures are shown in diagrams (Figure 2), but not in pseudocode format. |
| Open Source Code | No | The paper does not provide any statement or link indicating that the source code for the methodology is openly available. |
| Open Datasets | Yes | The efficacy of the proposed approach is demonstrated through rigorous ablation studies and exhaustive performance comparison on the popular KITTI outdoor driving dataset [Geiger et al., 2013]. |
| Dataset Splits | Yes | The KITTI dataset is divided into two splits namely, KITTI-Stereo split and KITTI-Eigen split which are commonly used to benchmark the performance of algorithms for depth and pose estimation. More details about the dataset and its use can be found in [Babu et al., 2018], [Godard et al., 2017] which are being omitted here for the sake of brevity. The depth evaluation is performed on both stereo and Eigen splits. |
| Hardware Specification | Yes | The total number of trainable parameters in this model is around 12 million and the network is trained for 240k iterations on Dell Alienware laptop having NVIDIA GTX 1080 GPU with 8 GB of GPU memory. |
| Software Dependencies | No | The paper states 'The proposed method is implemented in Tensorflow architecture.' but does not provide a specific version number for TensorFlow or any other software dependencies with version numbers. |
| Experiment Setup | Yes | The learning rate for training is initially set to 0.0001, then it is reduced by half after 3/5th of the iterations and further reduced by half after 4/5th of iterations. The γ value of the Charbonnier penalty [Babu et al., 2018] is set to 0.45. The α value in appearance loss is set to 0.85. The appearance loss and left-right consistency loss weights βap and βlr are set to 1.0. The smoothness loss weight βds is set to 0.1/s where s is the ratio of respective disparity image resolution to the input image resolution. The adversarial loss βg is set 0.001 which is obtained after extensive ablation study. |