End-to-End Line Drawing Vectorization
Authors: Hanyuan Liu, Chengze Li, Xueting Liu, Tien-Tsin Wong4559-4566
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Qualitative and quantitative evaluations show that our method achieves state of the art performance. We conduct comprehensive evaluations and ablation studies to validate the effectiveness of our method. We use a portion (excluding the BACKPACK and the BICYCLE classes for validation) of Quick Draw (Ha and Eck 2017) and the TU Berlin (Eitz, Hays, and Alexa 2012) dataset for the framework training. For the qualitative comparison, we first randomly sample drawings in the validation dataset. We implement our framework in Py Torch following a similar Transformer design as (Carion et al. 2020; Egiazarian et al. 2020) with 8 decoder layers for Stroke Encoder and 6 encoder/decoder layers for Stroke Vectorizer. The statistics are shown in Table 1. To verify the effectiveness of our framework design, we conduct ablation studies on different loss terms. |
| Researcher Affiliation | Academia | Hanyuan Liu,1 Chengze Li,2 Xueting Liu,2 Tien-Tsin Wong1* 1 The Chinese University of Hong Kong 2 Caritas Institute of Higher Education |
| Pseudocode | Yes | Algorithm 1: Stroke Vectorizer |
| Open Source Code | No | The paper states, 'More technical information, such as the detailed network architecture and the interconnect between each component, will be included in the supplementary material.' This is not a concrete statement that the source code will be released or is available. |
| Open Datasets | Yes | We use a portion (excluding the BACKPACK and the BICYCLE classes for validation) of Quick Draw (Ha and Eck 2017) and the TU Berlin (Eitz, Hays, and Alexa 2012) dataset for the framework training. |
| Dataset Splits | No | The paper mentions using a 'portion' of the datasets for training and specific classes for validation and testing, and provides the number of images (1150) for the testing dataset. However, it does not provide specific percentages or counts for the training and validation splits to allow full reproduction of the data partitioning. |
| Hardware Specification | Yes | We trained our model using 4 NVIDIA Titan V GPUs with automatic mixed precision, gradient accumulation trick, and gradient checkpointing (Chen et al. 2016) through all our experiments. |
| Software Dependencies | No | The paper mentions implementing the framework in 'Py Torch' but does not specify the version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | We implement our framework in Py Torch following a similar Transformer design as (Carion et al. 2020; Egiazarian et al. 2020) with 8 decoder layers for Stroke Encoder and 6 encoder/decoder layers for Stroke Vectorizer. Both of the Transformers use 8 attention heads. For Stroke Encoder, the 1D sinusoidal positional encoding is used for parallel decoding (Carion et al. 2020). We use three independent Adam optimizers for Stroke Encoder, Stroke Decoder, and Stroke Vectorizer, all with an initial learning rate of 0.0001. |