Klawens/Turtle

Turtle

Install:

cd mmdetection

pip install torchvision

pip install mmcv-full

pip install -r requirements/build.txt

pip install -v -e .

Convert the images to coco dataset format first

Train:

python tools/train.py own_configs/r50.py --no-validate

Inference and generate json result (other format please reference mmdetection official tutorial):

python tools/test.py own_configs/r50.py best_model.pth --format-only --eval-options "jsonfile_prefix=xxx"

The inference.py is for visualization.

All the parameters and dataset path, including IoU thres, confident thres, you can tune it in the own_configs/r50.py ( test_cfg )

Dataset:

Check data/make_dataset to convert the raw images to coco format for inference (get the boxes)

First using image2csv.py to generate a csv file with fake boxes. (mmdetection read coco format as default, so we can make a random labeled coco format test json, this csv file is a middle procedure.)

Then use the csv2coco.py to generate the fake test json. You only need to modify the classname_to_id at the top, and the path at the bottom of the script.

If you want to make a training set and a validation set, use the csv2coco_train_val.py, the real labeled csv file is required.

The scripts may occur type error depending on the image names, it won't be hard to fix.