shawn3298317/cv_practical

Repo for implementation of a circle-object detector which predicts the center & diameter of the circle object in the given img

Circle Object Detection

Development Env

Python 3.6

• numpy==1.16.4
• tensorflow==1.13.1
• scikit-image==0.16.2
• matplotlib==2.2.2
• shapely==1.6.4.post

Run Experiement

To evaluate with pre-trained model, simply run:

python main.py

To train model:

python train_model.py

(The log loss history will be written to output.txt)

Model Description

• Preprocess:
  1. Min-max normalization (divide by 3)
  2. Thresholding (0.7)
  3. Subsampling (2x2, max-sampling)
• Train-eval split:
  1. Small: 800 train, 200 eval
  2. Medium: 8000 train, 2000 eval
  3. Large: 40000 train, 10000 eval
• Network Topology (defined in train_model.py): Conv_3*10 + Conv_3*10 + Conv_2*10 + Conv_2*10 + Dense(50) + Dense(20) + Dense(10)
• Loss Function: mean squared error (performed better than huber loss)
• Optimizer: ADAM (lr = 0.001)
• Regularization: None (due to time issue)
• Trained for 150 epochs, batch size = 50
• Pretrained model checkpoint: model-0150.ckpt

Evaluation

• (Averge over 1000 samples w/t noise level = 2) IOU@0.7: 0.9673 (std_dev = 0.00642 for 6 experiments)

Demonstration

Yellow line is the true label and red line is the prediction from the cnn network.