HW2 ― Object Detection

Model: Yolov1

Demonstration

Using naive model

Using improve model

How to use

1. Download dataset

   bash get_dataset.sh
   

2. Download pretrained model and inference

   bash hw2.sh
   

3. Download pretrained improved model and inference

   bash hw2_best.sh
   

Requirements

Because I have reconstructed the project using tensorboard (pytorch-1.4.0~), so it's not compatiable in older version.

More Information

Please read requirement to get more information about this HW.

Performance Report

Question 1: Architecture and Hyperparameters

Model structure

Input size: (3, 448, 448)
Layer: 
    VGG16-bn: (512, 7, 7) 
    Flatten-layer: (25088) 
    Fully Connected: (4096) 
    LeakyReLU(0.02): (4096) 
    Dropout(0.5): (4096) 
    Fully Connected: (1274) 
    Sigmoid: (1274) 
    Reshape: (7, 7, 26)

Hyperparameters

• Optimizer: SGD
• Batch Size: 16
• Momentum: 0.9
• Weight Decay: 1e-4
• Learning Rate:
  • (1-20 Epoches) 1e-3
  • (21-40 Epoches) 1e-4
  • (41-55 Epoches) 1e-5
  • (56+ Epoches) 1e-6
• Non-maximum supression
  • Keep bounding box: 0.05 (Keep when P(class) > 0.05)
  • IOU threshold: 0.5 (Remove when IOU > 0.5)

Augmentation:

• Random Horizon Flip
• Random Zoom In (x1.1) and crop

Final Result:

Question 3: Improve Architecture and Hyperparameters

Model structure

Input size: (3, 448, 448)
Layer: 
    VGG16-without Maxpooling: (512, 14, 14)
    Conv2D(size=1) (26, 14, 14) 
    Sigmoid: (5096) 
    Reshape: (14, 14, 26)

Hyperparameters

• Optimizer: SGD
• Batch Size: 16
• Momentum: 0.9
• Weight Decay: 1e-4
• Learning Rate: 1e-3
• Non-maximum supression
  • Keep bounding box: 0.05 (Keep when P(class) > 0.05)
  • IOU threshold: 0.5 (Remove when IOU > 0.5)

Augmentation:

• Random Horizon Flip
• Random Vertical Flip

Final Result:

Question 5: mAP score

Model	mAP score	IoU	$P_{min}$
basic	10.55%	0.5	0.05
improve	11.76%	0.5	0.05

Question 6: Statistic of Dataset

Category	Final AP (Base)	Final AP (Improve)	Number in train15000
Plane	21.01%	35.75%	8723
Baseball-diamond	16.67%	15.58%	515
Bridge	1.14%	0.56%	2114
Ground-track-field	9.09%	0.0%	621
Small-vehicle	0.38%	9.09%	116228
Large-vehicle	9.09%	18.26%	23746
Ship	2.7%	14.27%	34585
Tennis-Court	36.48%	52.36%	3279
Basketball-court	14.54%	18.18%	661
Storage-tank	9.09%	10.08%	5199
Soccer-ball-field	32.32%	0.0%	590
Roundabout	0.0%	1.13%	537
Harbor	7.20%	2.28%	7457
Swimming-pool	9.09%	10.56%	1977
Helicopter	0.0%	0.0%	434
Container-crane	0.0%	0.0%	136