YoloV3 to track Alex

---

YoloV3 Simplified for training on Colab with custom dataset.

A Collage of Training images

We have added a very 'smal' Coco sample imageset in the folder called smalcoco. This is to make sure you can run it without issues on Colab.

Full credit goes to this, and if you are looking for much more detailed explainiation and features, please refer to the original source.

Thank you, The School Of AI!

Checking whether YoloV3 trains properly in colab.

1. Create a folder called weights in the root (YoloV3) folder
2. Download from: https://drive.google.com/open?id=1LezFG5g3BCW6iYaV89B2i64cqEUZD7e0
3. Place 'yolov3-spp-ultralytics.pt' file in the weights folder:

• to save time, move the file from the above link to your GDrive
• then drag and drop from your GDrive opened in Colab to weights folder

4. Open python 3 notebook in google colab.
5. Clone this repository using the below command:

!git clone https://github.com/genigarus/YoloV3.git

6. Run this command

python train.py --data data/smalcoco/smalcoco.data --batch 10 --cache --epochs 25 --nosave

If training occurs properly, the next stop is training with custom dataset.

For custom dataset:

Dataset Creation

Download images

I downloaded 500 images of Alex of Madagascar which is not present in COCO dataset.

Note:

1. Ensure proper names of images.
2. Please ensure that the images are within 800x800 so that it fits properly in the annotation tool window.

Annotate images

1. Clone this repo: https://github.com/miki998/YoloV3_Annotation_Tool
2. Follow the installation steps as mentioned in the repo.
3. Annotate the images using the Annotation tool.
4. This will create a separate label file for each image with image's file name within the Labels folder of YoloV3 Annotation tool.

Create dataset with structure as accepted by YoloV3

1. Place all the images in images folder and all the label files from above annotation tool in the labels folder.
2. Place both the images and labels folder within another folder. I have kept them in customdata folder.
3. Below is the folder structure which needs to be there:

data
  --customdata
    --images/
      --img001.jpg
      --img002.jpg
      --...
    --labels/
      --img001.txt
      --img002.txt
      --...
    custom.data #data file
    custom.names #your class names
    custom.txt #list of name of the images you want your network to be trained on. Currently we are using same file for test/train

4. As you can see above you need to create custom.data file. For 1 class example, your file will look like this:

  classes=1
  train=data/customdata/custom.txt
  test=data/customdata/custom.txt 
  names=data/customdata/custom.names

5. Below is the structure for custom.txt file. This is how the file looks like (please note the .s and /s):

./data/customdata/images/img001.jpg
./data/customdata/images/img002.jpg
./data/customdata/images/img003.jpg
...

6. You need to add custom.names file as you can see above. Since I downloaded images of Alex. My custom.names file looks like this:

alex

7. Alex above will have a class index of 0.

Architecture change

1. For COCO's 80 classes, YOLOv3's output vector has 255 dimensions ((4+1+80) * 3). Now we have 1 class, so we would need to change it's architecture.
2. Copy the contents of 'yolov3-spp.cfg' file to a new file called 'yolov3-custom.cfg' file in the data/cfg folder. In this file, make below changes:

i) Search for 'filters=255' (you should get three entries). Change 255 to 18 = (4+1+1) * 3 ii) Search for 'classes=80' and change all three entries to 'classes=1'. iii) For few samples, it is a good idea to change:

• burn_in to 100
• max_batches to 5000
• steps to 4000,4500

Training and detection Yolo v3 on your custom dataset

1. Follow this to train and detect your class.
2. Don't forget to perform the weight file steps mentioned in the section above.
3. Run this command python train.py --data data/customdata/custom.data --batch 10 --cache --cfg cfg/yolov3-custom.cfg --epochs 3 --nosave

As you can see in the collage image above, a lot is going on, and if you are creating a set of say 500 images, you'd get a bonanza of images via default augmentations being performed.

Results

After training for 300 Epochs, results look awesome!

Alex Detection Video