Tomato Allergies

Results with YOLOv3

Evaluation specs

Data specs

Dataset is shuffled and split in 2100/450/450 (train/val/test).

Then the images containing tomato bbox (394 such images) in training are upsampled, and final training size is 4061 (1706 without tomatoes + 2355 with tomatoes images).

Yolo specs for training

batch=32 ; subdivisions=16 ; width=640 ; height=640 ; momentum=0.9 ; decay=0.0005 ; learning_rate=0.001

Transfer learning

weights are initialized with darknet53 model trained on Imagenet.

wget https://pjreddie.com/media/files/darknet53.conv.74 to get it.

Error rate

Error rate on test set = 0.11 with checkpoint in release checkpoint_0.3.

The detection threshold used is 0.15.

Note: Overfitting is not fully attained and better ER is possible with more iterations.

Requirements installation

Simply python3 -m pip install . or python3 -m pip install -r requirements.txt

You need OpenCV >= 2.4 for compiling darknet with opencv2 and be able to compute metrics. (https://docs.opencv.org/3.4/d7/d9f/tutorial_linux_install.html). Check then /usr/include/opencv2.

Usage

Formatting

All images should be by default in ./data/assignment_imgs

img_annotations.json should be in ./data

label_mapping.csv should be also in ./data

or you can pass your own paths with --data-dir-path ,

--data-annotations-file_path ,

--labels-mapping-file-path

First transform the data in darknet format:

python3 -m tomato_dataset_tool

You could also upsample tomatoes for your training. Simply run instead:

python3 -m tomato_dataset_tool --upsample

All new files are located in ./data/formated (including gt text files for classification, i.e. if an image contains a tomato)

Train and test

Once it's done, Install darknet:

python3 -m darknet_handler --install

Note: you can pass --gpu to compile with gpu, also you can pass --opencv2 to compile with opencv (need for testing and computing metrics)

Training:

python3 -m darknet_handler --train --ckpts-file-path <path>

Testing and look at error rate:

python3 -m darknet_handler --test --ckpts-file-path <path>

It will create preds.txt in darknet-master directory. Then use compute_metrics.py to get the error rate:

python3 -m compute_metrics --yolo-output-filepath ./darknet-master/preds.txt --gt-filepath ./data/formated/test_gt.txt