- This repository implements FCOS(Fully Convolutional Ones Stage object detection) on the given dataset. Details about preprocessing and model architecture explanation has been discussed over here.
- Model which gave the best mAP score on test set was chosen. Please refer the documentation below to quickly start with data preparation, training and inference.
- Current implementation is based on Adelaidet framework which is based on detectron2.
- Tested CPU version successfully on Ubuntu 16.04 and GPU version on Ubuntu 18.04(Colab Version).
- Please make sure that python version is 3.7(conda env is recommended).Clone this repository and install requirements as shown below:
git clone https://github.com/rahulmangalampalli/FMCG-Fast-Moving-Consumer-Goods-Detection.git
pip3 install -r requirements.txt
pip3 install 'git+https://github.com/facebookresearch/detectron2.git' - We'll call this repository folder as the ROOT folder.
-
This repository supports coco format datasets. Download dataset and annotations from here and keep them in ROOT folder.
-
After downloading the data, extract files from it. It's structure is shown below:
|--ROOT
|--ShelfImages
|--train(283 images)
|--test(71 images)
|--annotations.csv- Run data_prep.py as shown below:
python3 data_prep.py --test_dir ShelfImages/test --train_dir ShelfImages/train --csv_file annotations.csv- This will give out two files, namely instances_train.json and instances_test.json which are in coco format.
-
Step 2 can be skipped if necessary, since instances_train.json and instances_test.json are already provided.
-
Install adelaidet as given below:
git clone https://github.com/aim-uofa/AdelaiDet.git
cd AdelaiDet
python3 setup.py build develop- Train Model using the command below:
python3 train.py --cpu #For cpu training, remove "--cpu" to switch to gpu- A colab version of training with gpu can be found here.(Recommended)
-
Download model file from here and keep it in the ROOT folder.
-
Inference on cpu:
python3 inference.py --im /path/to/image