FewSOLDataLoader 0.0.14

Pytorch Dataloader for FewSOL

FewSOL-DataLoader

This repo hosts the PyTorch dataloader for FewSOL dataset.

Using package

First install the package using

pip install FewSOLDataLoader

Setup

Step-1. Download the FewSOL dataset from https://irvlutd.github.io/FewSOL/#data

• There are four splits of the FewSOL dataset:
  1. real_objects : This is a real single object image split. Each object was captured from 9 angles
  2. real_clutter : This is a real clutter image split extracted from the OCID
  3. synthetic_objects : This is a synthetic single object image split made with 3D google objects. Each object was captured from 9 angles
  4. google_clutter : This is a synthetic clutter image split made with 3D google objects
     • Note: The google_clutter dataloader may take ~60 seconds to instantiate
• Note: The synthetic portion of the dataset is created using Google 3D Scanned Objects dataset.

Step-2. Pass the extracted dataset directory path into the dataloader as shown in the following example

Usage

Example

import random
from FewSOLDataLoader import load_fewsol_dataloader

 # Define the root directory
ROOT_DIR = os.getcwd()

# Define the dataset root directory using the join_path function
DATASET_ROOT_DIR = os.path.join(ROOT_DIR, 'FewSOL', 'data')
     
data = load_fewsol_dataloader(DATASET_ROOT_DIR, split="real_objects")    

# Generate a random index within the range of the dataloader's length
rand_idx = random.randint(0, len(data) - 1)

# Retrieve data from the dataloader for the random index
image_data, mask_data, bbox_data, label, questionnaire, file_name, poses = data[rand_idx]

# Synthetic objects and Real objects split also has a depth functionality
if s in ['synthetic_objects','real_objects']:
    depth = test.get_depth(rand_idx)
    print("Depth shape:", depth.shape)

Loading Specfic Data in order to speed up the dataloader

# Retrieve data from the dataloader for the random index
# Default loads all data, Data not loaded will be None
image_data, mask_data, bbox_data, label, questionnaire, file_name, poses = data.get_idx(
    rand_idx,
    load_img=False,
    load_mask=True,
    load_bbox=True,
    load_label=False,
    load_que=False,
    load_pose=False,
)

Getting indexs for a specfic class

# Gets the list of indexs for that contains a specific class
class_idxs = data.get_class_idx("bowl")
rand_class_idx = class_idxs[random.randint(0, len(class_idxs) - 1)]

Crop desired object using bbox data

# Functions supports 3D(color images) and 2D(no rgb axis)
from FewSOLDataLoader.helpers import crop_obj_using_bbox
rand_obj_idx = random.randint(0, len(label) - 1)
cropped_img = crop_obj_using_bbox(image_data[0],  bbox_data[0, rand_obj_idx])

Data Formats

• Image Data Shape

  # n x c x w x h
  # n = Number of total images
  # c = Number of Channels (RGB)
  # w = Width of the Image
  # h = Height of the image
  

• Semantic Segmentation Shape

  # n x m x w x h
  # n = Number of total images
  # m = Total number of objects in the current images
  # w = Width of the Image
  # h = Height of the image
  

• Detection Bounds Shape

  # n x m x r
  # n = Number of total images
  # m = Total number of objects in the current images
  # r = 4 : x, y, width, height
  

• Pose Information

  # n x m x 4 x 4
  # n = Number of total images
  # m = Total number of objects in the current images
  

• Label Output/Description Shape

  # m = Total number of objects in the images
  

Licenses

All files are licensed under the MIT license except for the below two inside FewSOL-DataLoader/src/FewSOLDataLoader/

• SingleRealPose.py - licensed under the NVIDIA Source Code License - Non-commercial as found here.
• CocoFormatConverter.py - licensed under the CC BY 4.0 LEGAL CODE as found here.

Bibtex

Please cite FewSOL if it helps your research:

@INPROCEEDINGS{padalunkal2023fewsol,
  title={FewSOL: A Dataset for Few-Shot Object Learning in Robotic Environments}, 
  author={P, Jishnu Jaykumar and Chao, Yu-Wei and Xiang, Yu},
  booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)}, 
  doi={10.1109/ICRA48891.2023.10161143},
  pages={9140-9146},
  year={2023}
}