Deep Learning - HW6: Object Detection

This is assignment 6 of Deep Learning, a course at Institute of Data Science, National Cheng Kung University. This project demonstrates the work of participating Find Mosquito Breeding Sources: Object Detection of Water Containers, a competition on AIdea with the team name jayliao. We utilized techniques of computer vision and image processing to detect 13 categories of water containers in images. Approach such as resizing, normalization, and non-maximum suppression were used to augment data. We adopted the structure of the single shot detector to detect targeted objects. Experiment showed that we achieve an acceptable performance (i.e., mAP: 0.4882 on the local testing set and 0.4542 on the submitted testing set) and indicated that stochastic gradient descent (SGD) was more appropriate than Adaptive Moment Estimation (Adam) to be the optimizer here.

Codes

• main.py: the main program for training and evaluation
• predict.py: the main program for prediction
• args.py defines the argument parser.
• trainer.py defines the trainer.
• model.py defines the model.
• datasets.py defines the class of training dataset with the torch.nn module.
• utils.py: tools for data loading

Data

Please go to the page of the competition on the AIdea here to download the following data.

• ./train_cdc/train_images/ contains 2,671 training images with the format of JPG.
• ./test_cdc/test_images/ contains 2,248 testing images with the format of JPG.
• ./train_cdc/train_annotations/: XML files of annotations for images under ./train_cdc/train_images/ with the format of PASCAL VOC. An XML file contains filename, size, and object of a single image. Each object has the name and (xmin, ymin) and (xmax, ymax) for the bounding box.

Please run utils.py with python3 utils.pycommand to obtain the following files:

• ./data_tables/table_tr.txt
• ./data_tables/table_va.txt
• ./data_tables/table_te.txt

After running main.py, the following files would be put under ./data_tables/.

• ./data_tables/label_map.json
• ./data_tables/images_tr.json
• ./data_tables/images_va.json
• ./data_tables/images_te.json
• ./data_tables/objects_tr.json
• ./data_tables/objects_va.json
• ./data_tables/objects_te.json

Usage

Training and evaluation

• To train a new model

The device can be set by using -d. The no. of epochs can be set as -e. For more details about arguments, please go to args.py.

python3 main.py -d 'cuda:1' -e 1000

• To train from a checkpoint

python3 main.py -d 'cuda:1' -dt '09-11-39-53__bs=8_epochs=500/' -cp 'checkpoint_ssd300.pth.tar' -e 500

• The outputs (checkpoints, training plots, predition output file, and annotated images) would be put under the subfolder with name of the time when the training started under ./output/.

Prediction

• To get the prediction csv file

The hyper-parameter min_score used in the prediction can be adjusted with -ms command.

python3 predict.py -d 'cuda' -dt '08-23-02-56__bs=8_epochs=2000/' -cp 'checkpoint_ssd300.pth.tar' -ms .01

• To get several annotated images (images from ./test_cdc/test_images/)

python3 predict.py -d 'cuda' -dt '08-23-02-56__bs=8_epochs=2000/' -cp 'checkpoint_ssd300.pth.tar' -detect

Examples of annotated images

Reference

1. Erhan, D., Szegedy, C., Toshev, A., & Anguelov, D. (2014). Scalable object detection using deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2147-2154).
2. https://github.com/sgrvinod/a-PyTorch-Tutorial-to-Object-Detection