Object Detection using Low-light Images

This project implements a Faster R-CNN model using TensorFlow's Object Detection API to identify objects in images taken from the ExDark dataset.

Training the model

Structuring the data

The original structure of the dataset is modified to look like this:

/low-light-detection
├── data
│   ├── annotations
│   ├── test
│   ├── train

where annotations contains all the annotation text files, test contains ~20% of the images from each of the 12 classes, and train contains ~80% of the images from each class.

Converting annotations to CSV

The annotations are converted to CSV using txt_to_csv.py so that they can be used to generate the required TFRecord files. From within the repo folder (/low-light-detection), run:

python txt_to_csv.py train

python txt_to_csv.py test

The generated CSV files should then be moved to data.

Generating TFRecord files

Next, run the following:

python generate_tfrecord.py --csv_input=data/train_labels.csv --image_dir=data/train --output_path=train.record

python generate_tfrecord.py --csv_input=data/test_labels.csv --image_dir=data/test --output_path=test.record

This will generate the record files needed for training the model.

Running the Jupyter notebook

The record files and labelmap.pbtxt need to be made available for training before executing the notebook. The method here involves uploading them to Google Drive and loading them from there.

Specific instructions and explanations about the code can be found in the notebook.

Running the detector

To simply test the detector locally, download the trained model from here. Then, execute the Jupyter notebook from the "Run inference" section after making the following changes:

• Remove the first "%cd" line
• Modify "from utils" to "from object_detection.utils", if necessary
• Set the correct paths for the variables PATH_TO_CKPT, PATH_TO_LABELS and PATH_TO_TEST_IMAGES_DIR
• Ensure that the images in PATH_TO_TEST_IMAGES_DIR are of the format image1.jpg, image2.jpg, and so on
• Modify the loop in TEST_IMAGE_PATHS if there are less/more than 4 images

Dependencies and configurations

Local

• Python: 3.5.4 (64-bit)
• numpy: 1.17.1
• Pillow: 6.1.0
• pandas: 0.25.1
• tensorflow: 1.14.0
• object-detection: 0.1
• jupyterlab: 1.1.1

Google Colab

• Runtime type: Python 3
• Hardware accelerator: GPU

References

1. Exclusively Dark (ExDark) Image Dataset
2. Creating your own object detector
3. How to train your own Object Detector with TensorFlow’s Object Detector API
4. Object Detection in Google Colab with Custom Dataset