I tested with the yolo model, and the results is amazing. Yolo could detect every rotation and distance except vertical carrots upside down. I rotated carrots from various distance, and Yolo could not detect carrots vertically upside down. Therefore, I tested with the testset vertically upside down. Sometimes, Yolo could detect carrots but sometimes not. So today, I will train carrots with vertically upside down position. I newly made carrot dataset upside down version, and used labelImg to generate xml files. I started to train carrots with total 542 datas 500 epochs. I guess it takes 2-3 days to complete … 2019.10.09 How to train YOLO with my own dataset? – 13 : Train for the carrots in vertical position of upside down 더보기

As the trained model could mostly detect horizontal carrots, however, it could not detect vertical carrots at all. Therefore, I took pictures of vertical carrot from various distance, and prepared for dataset to train models of vertical carrots. I am going to train yolo with new dataset after the training going on now ends. I guess the number of epochs of training is quite important since the loss of the training model becomes around 0.38-0.44 after 348 epochs compared to 0.7-0.8 after 100 epochs. So I will set the environment at least 500 epochs for each training. The loss becomes … 2019.10.08 How to train YOLO with my own dataset? – 12 : Training with new dataset including carrots of all position 더보기

I suceeded to train carrots located far away from camera. I found the problem for the last training that pictures are rotated 90 degrees and does not fit to xml file. Yesterday, I retrained yolo with dataset from various distance, and I tested today whether yolo could detect carrot which is far away from camera lens. Above pictures show that yolo could detect carrot from far away. Below video is the result of the test with webcam. However, the model could not detect carrots in vertical. Therefore I have to train vertical carrots to the model today. 2019.10.08 How to train YOLO with my own dataset? – 11 : Success to detect carrots from far distance 더보기

I want to train model to increase accuracy detecting carrots far away. Since recent training dataset depends both on distance and rotation, I decided to separate training steps – distance, rotation. I took pictures – 117 images – of carrot in various distance, and use labelImg to prepare datset for training. I started train from the accurately detecting carrot model, from loss 2.5. 2019.10.07 How to train YOLO with my own dataset? – 10 : Prepare for new dataset and train model. 더보기

After 300 epochs of training, and the loss of the training is around 0.7-0.8 which is quite high. , I tested with the testset I made at first. The following pictures are the results of the test. I found the reason why the model is not trained well. In the folder, the pictures are rotated 90 degree to the left. However, since the xml file is not rotated too, I guess the results of the training becomes not accurate. 2019.10.07 2019.10.04 How to train YOLO with my own dataset? – 9 : Test for the trained model of distance. 더보기

I tested my yolo trained model with webcam whether it detects carrots spontanuously. Following videos are testing video. In the video, the accuracy decreased as the distance between webcam and carrots increases. Therefore, I need to train yolo with carrot pictures from various distance. I took a carrot pictures up and down. I obtained 135 images to train. I started to train yolo with newly provided datas. 2019.10.04 How to train YOLO with my own dataset? – 9 : Increase the accuracy of detecting carrots according to distance 더보기

I finally succeed to detect my own carrot! I used yolo.cfg file and yolo.weights file. The loss decreases around 0.3-0.4 as same as the soccer training, and the results is amazing. I finally suceeded to detect my lovely own carrot images. I also tested with my web camera, and it seems like the accuracy of yolo depends on the distance of the object. Since I trained the model with similar sizes of carrot pictures, yolo could not detect carrot far from camera correctly. 2019.10.03 How to train YOLO with my own dataset? – 8 : Succeed to detect my own carrot! 더보기

Following the instructions of the below website, I downloaded all the dataset and trained yolo with soccerball today since I want to test whether other datasets could train yolo good or not. I trained yolo, and the loss converges around 0.3 at 65 epoch.I tested with the sample images, and surprisingly yolo could detect soccer ball as well!!! I was so surprised at this result, and try to find the reasons why my dataset could not detect carrots so far. Numbers of Datasets (carrots – 100, ball – 195) Resolution of the imgs (carrots – 1920 x 1090 , ball … 2019.10.02 How to train YOLO with my own dataset? – 7 : Exercise with other examples 더보기

As I trained yolo with my own carrots, however, yolo still could not detect carrot correctly. I trained 300 epochs, and the loss decreases from 100 to 0.4. I adjusted threshold as 0.013, and tested with testset. However, the detection could not be done really good. So I want to try other datasets from other people, and trained yolo again. 2019.10.02 How to train YOLO with my own dataset? – 6 : Adjusting threshold 더보기