Neural Inverse Knitting: From Images to Manufacturing Instructions
Authors: Alexandre Kaspar, Tae-Hyun Oh, Liane Makatura, Petr Kellnhofer, Wojciech Matusik
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present our deep neural network model that infers a 2D knitting instruction map from an image of patterns. In this section, we provide the theoretical motivation of our framework, and then we describe the loss functions we used, as well as implementation details. ... We train our network with a combination of the real knitted patterns and the rendered images. We have oversampled the real data to achieve 1:1 mix ratio with several data augmentation strategies, which can be found in the supplementary material. We train with 80% of the real data, withholding 5% for validation and 15% for testing, whereas we use all the synthetic data for training. ... Table 1 compares the measured accuracy of predicted instructions on our real image test set. We also provide qualitative results in Figure 9. |
| Researcher Affiliation | Academia | 1Computer Science & Artiļ¬cial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology (MIT), Cambridge, MA, USA. |
| Pseudocode | No | The paper describes algorithms and procedures in text and through architectural diagrams (e.g., Figure 1 implies a structure but is not a pseudocode block), but does not provide explicit pseudocode or algorithm blocks. |
| Open Source Code | No | The paper states 'Project page: http://deepknitting.csail.mit.edu' but does not explicitly confirm that the source code for the methodology is provided there. This is a project page, not necessarily a code repository. |
| Open Datasets | Yes | To this end, we collect a paired dataset of knitting instruction maps and corresponding images of knitted patterns. We augment this dataset with synthetically generated pairs obtained using a knitting simulator (Shima Seiki). ... The main consideration for capturing knitted patterns is that their tension should be as regular as possible so that knitting units would align with corresponding pattern instructions. We initially proceeded with knitting and capturing patterns individually but this proved to not be scalable. We then chose to knit sets of 25 patterns over a 5 5 tile grid, each of which would be separated by both horizontal and vertical tubular knit structures. The tubular structures are designed to allow sliding 1/8 inch steel rods which we use to normalize the tension, as shown in Figure 7. Note that each knitted pattern effectively provides us with two full opposite patterns (the front side, and its back whose instructions can be directly mapped from the front ones). This doubles the size of our real knitted dataset to 2,088 samples after annotating and cropping the knitted samples. |
| Dataset Splits | Yes | We train with 80% of the real data, withholding 5% for validation and 15% for testing, whereas we use all the synthetic data for training. |
| Hardware Specification | Yes | The training took from 3 to 4 hours (depending on the model) on a Titan Xp GPU. |
| Software Dependencies | Yes | We implemented it using Tensor Flow (Abadi et al., 2016). ... We trained our model for 150k iterations with batch size 2 for each domain data using ADAM optimizer with initial learning rate 0.0005, exponential decay rate 0.3 every 50, 000 iterations. |
| Experiment Setup | Yes | We trained our network with a combination of the real knitted patterns and the rendered images. We have oversampled the real data to achieve 1:1 mix ratio with several data augmentation strategies, which can be found in the supplementary material. ... We trained our model for 150k iterations with batch size 2 for each domain data using ADAM optimizer with initial learning rate 0.0005, exponential decay rate 0.3 every 50, 000 iterations. |