CropRow Detection Lincoln Dataset (CRDLD)

Here we will release the version 2.1 of CRDLD dataset. Click here for version 1.0.
Download dataset

Dataset Structure

The dataset is based on 2000 images. The 2000 image dataset is classified into three groups for the purpose of training, testing and validation. The training, testing and validation sub-groups contain 1250,500 and 250 images respectively. Each image in the dataset consists of a corresponding ground truth image. Numeric values of the labelled coordinates are stored in .mat files.

.
│
├── Train Dataset              # 1250 images and ground truth segmentations from 50 data classes. [25 Images per class x 50 Classes]
|   └──labels.mat             # .mat file containing image label coordinates for 1250 images
│
├── Test Dataset               # 500 images and ground truth segmentations from 50 data classes. [10 Images per class x 50 Classes]
|    └──Labels                  # .mat file containing image label coordinates for 500 images
│
└── Validation Dataset               # 250 images and ground truth segmentations from CRDLDv1.0 base images.
    └──Labels                  # .mat file containing image label coordinates for 250 images

Table 1: Data Categories

Table 1 contains the different field variations represented in the dataset.

ID	Data Category	Description
a	Horizontal Shadow	Shadow falls perpendicular to the direction of the crop row
b	Front Shadow	Shadow of the robot falling on the image captured by the camera
c	Small Crops	Crop rows at early growth stages (Up to 4 unfolded leaves)
d	Large Crops	Presence of one or many largely grown crops (more than 4 unfolded leaves) within the crop row
e	Sparse Weed	Sparsely grown weed scattered between the crop rows
f	Dense Weed	Weed grown densely among the crop rows where the inter-row space is completely covered
g	Sunny	Crop row data captured in sunny weather
h	Cloudy	Crop row data captured in cloudy weather
i	Discontinuities	Missing plants in the crop row which lead to discontinuities in crop row
j	Slope/ Curve	Images captured while the crop row is not in a flat farmland or where crop rows are not straight lines
k	Tyre Tracks	Tyre tracks from tramlines running through the field

Figure 1: Example images from data categories in Table 1.

Table 2: Data Classes Formed by Combinations of Field Variations

Table 2 contains the combinations field variations given in Table 1 in the dataset.

Data Category	a	b	c	d	e	f	g	h	i	j
Horizontal Shadow	1
Front Shadow		7
Small Crops	2	8	11, 44		47				45	46
Large Crops	3	9	12	20, 48	50				49
Sparse Weed			13	21
Dense Weed			14	22
Sunny			15	23	28	32
Cloudy			16	24	29	33
Discontinuities	4	10	17	25	30	34	36	39
Slope/ Curve	5		18	26	31	35	37	40	42
Tyre Tracks	6		19	27			38	41	43

Sample Data

The crop row is labelled with white lines on black background. The line width of the white line is 6 pixels. The labels could be regenerated with custom line width using the labels .mat files.

Figure 2: Example image and semantic label from the dataset.

Citation

@article{de2022deep,
  title={Deep learning-based Crop Row Following for Infield Navigation of Agri-Robots},
  author={de Silva, Rajitha and Cielniak, Grzegorz and Wang, Gang and Gao, Junfeng},
  journal={Journal of Field Robotics},
  year={2023}
}