Neural Inverse Rendering for General Reflectance Photometric Stereo

Authors: Tatsunori Taniai, Takanori Maehara

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We present a novel convolutional neural network architecture for photometric stereo (Woodham, 1980), a problem of recovering 3D object surface normals from multiple images observed under varying illuminations. ... Our method is shown to achieve the state-of-the-art performance on a challenging real-world scene benchmark. In this section we evaluate our method using a challenging real-world scene benchmark called Di Li Gen T (Shi et al., 2018). In Sec. 4.1, we show comparisons with state-of-the-art photometric stereo methods. We then more analyze our network architecture in Sec. 4.2 and weak supervision technique in Sec. 4.3. In the experiments, we use M = 96 of observed images for each scene provided by the Di Li Gen T dataset.
Researcher Affiliation Academia 1RIKEN Center for Advanced Intelligence Project (RIKEN AIP), Nihonbashi, Tokyo, Japan.
Pseudocode No The paper describes the method and network architecture in detail with diagrams and text, but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code No The paper states, 'Our method is implemented in Chainer (Tokui et al., 2015)', but it does not provide any specific link or explicit statement regarding the open-sourcing of the code developed for this paper.
Open Datasets Yes In this section we evaluate our method using a challenging real-world scene benchmark called Di Li Gen T (Shi et al., 2018).
Dataset Splits No The paper describes unsupervised learning performed directly on individual test scenes without pre-training, stating, 'For each test scene, we iterate SGD updates for 1000 steps.' It does not specify traditional training/validation/test dataset splits.
Hardware Specification Yes Our method is implemented in Chainer (Tokui et al., 2015) and is run on a single n Vidia Tesla V100 GPU with 16 GB memory and 32 bit floating-point precision.
Software Dependencies No The paper mentions that the method is 'implemented in Chainer (Tokui et al., 2015)' and 'We use Adam (Kingma & Ba, 2015) as the optimizer', but no specific version numbers for these software components are provided.
Experiment Setup Yes For each test scene, we iterate SGD updates for 1000 steps. Adam s hyper-parameter α is set to α0 = 8 10 4 for first 900 iterations, and then decreased to α0/10 for last 100 iterations for fine-tuning. We use the default values for the other hyper-parameters. The convolution weights are randomly initialized by He initialization (He et al., 2015).