Database content
Jacquard dataset is composed of 54 485 different scenes from 11 619 distinct objects with a total of 4 967 454 grasps annotations (1 181 330 unique locations). For each scene, a rendered RGB image, a segmentation mask, two depth images and the grasps annotations are provided. More details about the creation of the dataset can be found in the paper.
Demonstration samples
An archive with a subset of 10 objects of the dataset can be dowloaded here (55.2 MB). To download the whole dataset, please follow the instructions on home page.
Files content
Each downloadable zip file contains an independant subpart of the dataset so you can work the 10 demo objects just as with the whole dataset. Once extracted, the structure of the dataset is the following:
- nameOfObject1: a folder containing all the data for the object. The name is consistent with ShapenetSem dataset. All the files inside this folder are prefixed with * = i_nameOfObject1, where i is the index of the scene (between 0 and 4) for the current object.
- *_grasps.txt: a text file with the grasps annotations. Each line in the file is one grasp written as x;y;theta in degrees;opening;jaws size. Please note that all values are in image coordinates, so they are expressed in pixels, y is toward the bottom of the image (and therefore the angle is horizontally mirrored). When the position is the same on multiple consecutive rows, the first one corresponds to the grasp with the default jaws size of 2 cm and the followings are just repetition of this grasp with different sizes.
- *_mask.png: a png image with a binary mask (0 for background) of the object which can be used to perform background augmentation on the data for example.
- *_perfect_depth.tiff: a float32 tiff image with the perfect depth image.
- *_RGB.png: a png image of the scene rendered with Blender.
- *_stereo_depth.tiff: a float32 tiff image with the depth obtained from an off-the-shelf stereovision algorithm. Pixels with an undetermined depth (because of occlusion or bad matching during the stereovision process) have a value of -1.
- nameOfObject2
- *_grasps.txt
- *_mask.png
- *_perfect_depth.tiff
- *_RGB.png
- *_stereo_depth.tiff
- nameOfObject3
- *_grasps.txt
- *_mask.png
- *_perfect_depth.tiff
- *_RGB.png
- *_stereo_depth.tiff
- ...