SSD: Single Shot MultiBox Object Detector

Installation

Follow this repository to set up the SSD PyTorch version on your system.

Dataset Preparation

PreProcessing.ipynb walks you through the dataset preparation steps.

Dataset Directory Strucure (PASCAL VOC 2007):

VOCdevkit/ VOC2007/ Annotations/ (.xml files of the 354 images) ImageSets/ Main/ train.txt (list of train image filenames) test.txt (list of test image filenames) JPEGImages/ (354 .JPG images)

Dataset Augmentation

Photometric Distortions
- Random Contrast
- Random Saturation
- Random Hue
- Random Brightness
- Random Lighting Noise
- Convert Colour Spaces
Expand
Random Sample Crop
Random Mirror

Detection Network, Package Versions and Pretrained Model

Detection Network: SSD300-VGG16 Package Versions: PyTorch 1.0.0, Torchvision 0.2.1, Pillow 6.1 Pretrained Model: https://s3.amazonaws.com/amdegroot-models/ssd300_mAP_77.43_v2.pth

Hyperparameters and Anchor Tuning

Batch Size: 32 Initial Learning Rate: 1e-3 Momentum: 0.9 Weight decay for SGD: 5e-4 Gamma update for SGD: 0.1

In the original code: Number of Anchor Boxes - 6 {Aspect Ratios: 1, 2, 3, 1/2, 1/3 and another 1:1 aspect ratio box} Total Number of Boxes - 38x38x4 + 19×19×6 + 10×10×6 + 5×5×6 + 3×3×4 + 1×1×4 = 8732

In this exercise: Number of Anchor Boxes - 1 {Aspect Ratio: 0.65} Total Number of Boxes - 38x38x1 + 19×19×1 + 10×10×1 + 5×5×1 + 3×3×1 + 1×1×1 = 1940

Average of all bounding box width and height is: (201.353, 308.988 =) This ratio is approximately 0.65. Hence, this was chosen as the aspect ratio of the single anchor box.