This project using RCNN (Region-CNN) for object detection.

The dataset is from CrowdAI, it contains 9,421 images with 72,064 entries of labels in three classes (car, truck, and pedestrain).

The pre-trained model is faster_rcnn_resnet101_coco available on TensorFlow website.

1. Set up the environment

• Install Ubuntu 18.04
• Install Nvidia driver 390 and Nvidia Docker 2
• Pull the tensorflow:1.12.0-gpu-py3 image and run it using runtime of nvidia
• Get into the container
• Install the TensorFlow Object Detection APIs and COCO APIs

2. Prepare the scripts needed

• Modify the generate_tfrecord.py to fit the needs of the file
• Construct label_map.pbtxt manually

3. Prepare tfrecord file using the dataset

• Download the dataset mentioned above and extract it using tar -xvzf
• Download the corrected .csv file
• Split the .csv file into train.csv and test.csv
• Used generate_tfrecord.py to convert the csv files into tfrecord

4. Retrieve pre-trained model

• Download the pre-trained model from TensorFlow website
• Extract

5. Config pipeline

• Modify the pipeline.config file in the model
• Set the classnames, tfrecords, and other parameters

6. Training

• Set environment parameters
• Run model_main.py with parameters to train
• Launch Tensorboard for monitoring

7. Inference on specific images

• Use the helper code from object_detection_tutorial.ipynb
• Set the path to test images and image parameters
• Show image with inferred labels

The model took 60 hours to train. After 50,000 steps of training, the Mean Average Precision (mAP) reached 0.3693. The mAP for large objects reached 0.5546.