Closed mhammoud-os closed 6 months ago
Found some useful links:
I just started with vertex AI by Google, I think I want to use it, however their jsonl format is quite different then the json format by the roboflow datasets. I am just finiched creating a python program that help me converting the formats.
The json format that I picked from roboflow is Create ML
Added 70 images from this dataset: https://universe.roboflow.com/garbage-time/weedbot-luupi/dataset/1
Added aprox 80 images from this dataset: https://universe.roboflow.com/treedetectiontest-bitpq/tree_detection_test/dataset/3
Added 500 imges from this dataset: https://universe.roboflow.com/yoshiki-terada/finalmodelforweeddetectioningrassland
Am going to remove them because I did'nt like the quality of images
Added aprox 50 images from this dataset: https://universe.roboflow.com/simon-blackmore-wftzr/smart-weeder/dataset/3
added aprox 100 images from this dataset: https://universe.roboflow.com/flower-ppbuo/dandelion-5kbfw/dataset/1
The last two datasets contained the exact same images. This may cause over fitting, I will remove the first one
Added Aprox 40 images from this dataset: https://universe.roboflow.com/clover-detection/clover_3/dataset/1
Added aprox 200 from this dataset: https://universe.roboflow.com/university-of-burgandy-zowkw/detection_plant/dataset/1
added aprox 200 from this dataset: https://universe.roboflow.com/spes-robotics/plaintain_rotated/dataset/1
I have little over 1000 images, My data has been collected. However I will need to go through all the images and refine this dataset. Good data = Good model
Finished refining dataset started training.(dataset is now 750 images total) I will probably have to adust the data even more(add or remove) but will have to wait and see. The model will do it's own tests by spliting the data into 3 groups but I want to do my own tests after with a few random images.
Model is done and exported but I needed to Migrate Raspberry Pi from SD to USB since I ran out of storage space to store tenserflow, open-cv, numpy, etc.
I had problems when migrating raspayr pi. The reason why I needed to do it is because I have 2 os's stored on the sd card, ubuntu server and rasbian, they are controlled using PINN os. I could not delete os's using pinn so I eventually found out I needed to move them using tar. Ubuntu server worked very well with tar, however rasbian did not. I eventually just disided to manually move my files and download the needed pacages again. Because I had a problem with downloading open-cv on debain 11, I used debain 12 this time(bookworm) It worked perfectlly. I was abel to put everything back together and run my object detection model. I did my test on a few random outside images, i am happy with the results, however I think I need to add more images for certain weeds like the clover. I will calculate the score later at the very end.
Precision: 89.1% Recall: 62.4%
This model should be able to detect all the common weeds found in yards in all stages of development. Steps: