Yes i know about this. this dataset created for carrot crop and weed using Bonirob robot. But in India, it is very difficult to get fund for such a machinery. i wanted to know whether images can be acquired under natural conditions and using a camera fixed on fixed distance from the ground?

Please elaborate the application you are interested in with these crop images. This will decide your benchmark for each image. For example, if your problem is classification of crops, then each image will be labelled as name of the crop seen in that image.

Thanks for your response. I am working on automated detection and classification of weeds in paddy field images. it will be base for computer vision application in agriculture (Precision Agriculture), for this I wanted to create dataset. Please help me out.

Development of agricultural database can be a extremely difficult task. Even after long efforts, your database may not give good results for test cases captured in different natural conditions. The main challenges in making the database are the large variety of any agricultural product and its disease, background, lighting conditions etc. So, if you want to make your database in natural conditions be careful. Though, there is no doubt that we want to achieve good detection/classification rate in natural environments, our present algorithms are not so robust. If you are working with machine learning algorithms, you may require domain adaptations or robust feature extractions to achieve decent results. 

If you do not want to take the challenge and want to create an ideal dataset, make sure you have constant illumination (preferably diffused as vegetables may have specular reflecting surface), static background (preferably flat color, so that you can segment the vegetable easily) and as many variations of the vegetable (normal and with weeds) as possible. You may or may not include multiview. 

thanks for your response. i have started creating dataset under variable lighting condition. what should be the accuracy in this type of environment according to you?

can i send some of the images and annotated ones so that you can give your valuable suggestions on it?

The accuracy will depend on the features that you will select, and on the samples that you are taking. If you are working with features which are robust to color and intensity changes, you will have decent accuracy. On the other hand, if your training set already have samples close to your testing conditions, your accuracy may again improve.

It is not possible to comment about accuracy without knowing the database, methods, imaging conditions etc. You can achieve good accuracy in varying illumination which makes the problem a lot more challenging ( and interesting) for researchers. I am really a beginner in this field. I think, you can share some of your images here so that the experts can give  their valuable comments. 

Thanks for your mail. i am sending the original images, images after soil and background masking and annotated images.(Sample dataset) images were captured under natural conditions in paddy fields using tripod fixed at a distance of 3 ft from ground. please give valuable input on this thank you .