Multispectral Transfer Network: Unsupervised Depth Estimation for All-Day Vision
Authors: Namil Kim, Yukyung Choi, Soonmin Hwang, In So Kweon
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In experiments, we demonstrate the better performance and generalization of depth estimation through the proposed multispectral stereo dataset, including various driving conditions. |
| Researcher Affiliation | Collaboration | Namil Kim, 1,2 Yukyung Choi, 1,3 Soonmin Hwang,1 In So Kweon1 1Korea Advanced Institute of Science and Technology (KAIST), Korea 2NAVER LABS Corp., Korea 3Clova, NAVER Corp., Korea |
| Pseudocode | No | The paper does not contain an explicit pseudocode block or a clearly labeled algorithm block. |
| Open Source Code | No | The paper provides a project website URL (http://multispectral.kaist.ac.kr) which states "The entire network configuration is illustrated in the website.", but it does not contain an unambiguous statement that the source code for the described methodology is publicly released or a direct link to a code repository. |
| Open Datasets | No | The paper introduces a new "large-scale multispectral stereo dataset" and describes its characteristics, but it does not provide concrete access information (e.g., a direct link, DOI, or formal citation with authors and year) for this dataset. |
| Dataset Splits | Yes | In total, we provide (#7383) stereo/thermal images for training (#4534) and testing (#2853) during the daytime, while also testing (#1583) pairs at night. We split the training/testing samples using GPS and time-logging data without unnecessary duplication or consistency issues. |
| Hardware Specification | Yes | The network is implemented in Mat Convnet (Vedaldi and Lenc 2015), and takes 20 hours to train using a single NVIDIA TITAN X GPU on 4.5 thousand pairs for 40 epochs. |
| Software Dependencies | No | The paper mentions "Mat Convnet" as the implementation framework but does not specify its version number or other software dependencies with their respective version numbers. |
| Experiment Setup | Yes | During the training process, we set the weights of the objective terms as λs = 0.01 and λc = 0.01 and use SGD for optimization with a momentum(0.9)/weight decay(0.0005) from scratch within a learning rate of 10-5. According to the normalized coordinate of the bilinear sampler, we set an adaptive scaled sigmoid function which initially sets β0 = 0.3, with an increase by α = 0.01 every two epochs. |