Emergence of Object Segmentation in Perturbed Generative Models
Authors: Adam Bielski, Paolo Favaro
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate this framework on real images of several object categories. To evaluate the quality of the generated segments, we generate 10,000 images and masks for each setting... The quantitative results can be found in Table 1 and the qualitative results in Fig. 4. Ablation study. To validate the design choices in our approach we perform ablation experiments on the LSUN Car dataset. |
| Researcher Affiliation | Academia | Adam Bielski University of Bern adam.bielski@inf.unibe.ch Paolo Favaro University of Bern paolo.favaro@inf.unibe.ch |
| Pseudocode | No | No structured pseudocode or algorithm blocks were found. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing its own source code, nor does it provide a link to a repository for the methodology described. |
| Open Datasets | Yes | We train our generative model on 4 LSUN object categories (Yu et al. [2015]): car, horse, chair, bird. We repeat the same training and evaluation procedure on Caltech-UCSD Birds-200-2011 dataset (Wah et al. [2011]), for which the segmentation ground truth is available. |
| Dataset Splits | No | The paper mentions training on datasets and evaluating performance, but does not provide specific details on validation dataset splits (e.g., percentages, sample counts, or explicit validation set usage). |
| Hardware Specification | No | No specific hardware details (like GPU models, CPU types, or cloud instance specifications) used for running experiments were mentioned. |
| Software Dependencies | No | The paper mentions software components like Style GAN, Mask R-CNN, and Adam optimizer but does not provide specific version numbers for these or other ancillary software. |
| Experiment Setup | Yes | We start the training with an initial resolution of 8 8 pixels and use progressive training to up to 128 128 pixels. We train with batch sizes 256, 128, 64, 32 and 32 for resolutions 8 8, 16 16, 32 32, 64 64 and 128 128 respectively. For each scale the number of iterations is set to process 1,200,000 real images. We set the discriminator loss parameters to λ = 10 and ϵ = 0.001. In the generator loss we set γ1 = 2 for the minimum mask size term and γ2 = 2 for the binarization term. We optimize our GAN with the Adam optimizer (Kingma and Ba [2015]) and parameters β1 = 0, β2 = 0.99. We use a fixed learning rate of 0.001 for all scales except for 128 128 pixels, where we use 0.0015. We run the training for 1000 iterations with Adam optimizer, learning rate of 0.0001 and β1 = 0.9, β2 = 0.999. |