Zero Pixel Directional Boundary by Vector Transform
Authors: Edoardo Mello Rella, Ajad Chhatkuli, Yun Liu, Ender Konukoglu, Luc Van Gool
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the proposed loss method using a standard architecture and show the excellent performance over other losses and representations on several datasets. |
| Researcher Affiliation | Academia | 1 Computer Vision Lab, ETH Zurich, Switzerland 2 VISICS, ESAT/PSI, KU Leuven, Belgium |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | The paper does not provide any statement or link regarding the availability of its source code. |
| Open Datasets | Yes | We compare the proposed method on Cityscapes (Cordts et al., 2016), Mapillary Vistas (Neuhold et al., 2017) and Synthia (Ros et al., 2016), three datasets providing high quality instance and semantic boundaries. Despite the inconsistent thickness of annotated boundaries, we also compare our method on BSDS500 Arbelaez et al. (2010). |
| Dataset Splits | Yes | Cityscapes train: 2500 validation: 475 test: 500 ... Synthia train: 6600 validation: 800 test: 1600 ... Mapillary Vistas train: 17000 validation: 1000 test: 2000 ... BSDS500 train: 200 validation: 100 test: 200 |
| Hardware Specification | No | The paper does not explicitly describe the hardware used to run its experiments. |
| Software Dependencies | No | The paper mentions architectural components and methods like Batch Normalization, ReLU, Sobel filters, and ResNeXt, but does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow, CUDA versions). |
| Experiment Setup | Yes | More specifically, we apply an initial learning rate of 0.001, the poly learning rate scheduler (Liu et al., 2015) and a batch size of 32. The images are augmented using random resizing, random horizontal flipping and randomly cropping the resulting image to the size of 512 × 512 irrespective of the dataset. The dimensions for randomly resizing the shortest side are from the set of dimensions {512, 640, 704, 832, 896, 1024, 1152, 1216, 1344, 1408, 1536, 1664, 1728, 1856, 1920, 2048}. |