Code, demos and data for SketchParse (a neural network for sketch segmentation). Paper:
This is a neural network for (semantic) sketch segmentation. Use it to associate semantics with your freehand sketches!
The four panels are chosen from the 100th, 75th, 50th, 25th percentile accuracy of segmentation (by IoU) respectively. As you can see, even in (relatively) bad cases, we can provide fairly accurate segmentations!
We have a multi-task deep neural network that can segment freehand sketches as well as predict a global pose:
For code and corresponding instructions, navigate to exp-src
For annotation software tool and corresponding instructions, navigate to valsketch
For retrieval code, models and a demo, navigate to retrieval-src
For instructions to get annotated sketch dataset, navigate to exp-src/data/sketch-dataset. Pose dataset is present in exp-src/data/lists/Pose_all_label.txt
. Find instruction regarding pose dataset here.
This code was developed and tested on an Ubuntu 14.04 machine with python 2.7 and pyTorch (v0.1.12). We used an NVIDIA TITAN X for training and evaluating our model.
SketchParse: Towards Rich Descriptions For Poorly Drawn Sketches Using Multi-Task Deep Networks
For questions regarding the main segmentation network, please contact Isht ([email protected])
For questions regarding the annotation tool, please contact Sahil ([email protected])
For questions regarding the pose subnetwork and the sketch-based image retrieval application, please contact Abhijat ([email protected])
For any other questions, please contact Ravi ([email protected])