SVHNDetection

This is a task of object detection with The Street View House Numbers (SVHN) Dataset. There are 10 number classes in the dataset. We use 30,000 images for training, 3,402 for validation, and 13,068 for testing. Use YOLOv5 model from GitHub and train with pretrained checkpoint. The highest testing mAP can reach 52.115%.

Reproducing Submission

1. Installation
2. File Configuration
3. Dataset Configuration
4. Training
5. Inference

Installation

1. Clone this repository.

git clone https://github.com/qaz812345/SVHNDetection

2. We will use YOLOv5 model from GitHub. cd to this repository and clone YOLOv5 repository.

cd SVHNDetection
git clone https://github.com/ultralytics/yolov5

3. Install the pytyhon packages with requirements.txt.

pip install -r requirements.txt

File Configuration

Please move files to the corresponding destination.

File Name	File Destination
parse_data.py	yolov5/
detect.py	yolov5/
datasets.py	yolov5/utils/
svhn.yaml	yolov5/data/

Dataset Configuration

To train SVHN dataset on YOLOv5, we need to set up the configuration for it.

1. Download train.tar.gz and test.tar.gz from SVHN dataset site.
2. Unzip the data files and set the data directory structure as:

SVHNDetection
|__svhn
|   |__images
|   |   |__train
|   |   |  |__<train_image>
|   |   |  |__ ...
|   |   |  
|   |   |__val
|   |   |  |__<val_image>
|   |   |  |__ ...
|   |   | 
|   |   |__test
|   |      |__<test_image>
|   |      |__ ...
|   |     
|   |__labels
|       |__train
|       |
|       |__val
|
|__yolov5
|   |__ ...
|
|__ parse_data.py

3. Pick a number as dividing index (ex. 30000) to devide training images into training set and vaildation set. Move the images after dividing index (ex. 30001) to svhn/images/val as vaildation set
4. Parse label data:

python parse_data.py -f ../svhn/images/train/digitStruct.mat -d <dividing_index>

Training

Model Configuration

1. Download yolov5l.pt weights from here.
2. Edit yolov5/models/yolov5l.yaml line 2 as:

# parameters
nc: 11  # in svhn 0 is labeled as '10' so we have 11 classes '0' to '10'. And class '0' will never appear.
...
...

Data Pre-process and Augmentation

• Random mosaic
• Resize image (shape=(640, 640))
• Random translation (translate=0.1)
• Random scale (scale=0.5)
• Random horizontal flip

You can edit yolov5/data/hyp.scratch.yaml for changing augmentation parameters or using the other augmentation methods.

Hyperparameters

• Epochs = 100
• Batch size = 32
• Optimizer = SGD (learning rate=0.01, momentum=0.937, weight decay=0.0005)
• Warmup (epochs=3, momentum=0.8, bias learning rate=0.1)
• Scheduler = cosine learning rate decay (final OneCycleLR learning rate=0.2)
• Loss function = IOU loss
• Box loss gain = 0.05
• Class loss gain = 0.5
• Object loss gain = 1.0
• Device = 0,1,2,3

You can edit yolov5/data/hyp.scratch.yaml for changing some hyperparameters and some need to specify at training time.

Train

Training command:

python train.py --weights weights/yolov5l.pt --cfg models/yolov5l.yaml --data data/svhn.yaml --epochs 100 --batch-size 32 --device 0,1,2,3

If your GPUs are out of memory, please decrease batch size or change to smaller model like yolov5m or yolov5s. The way of changing configuration setting is simlar to yolov5l, please check Model Configuration section above.

Inference

Output structure of json file:

[{"bbox": [(y1, x1, y2, x2), (y1, x1, y2, x2)], "score": [conf, conf], "label": [cls, cls]},
{"bbox": [], "score": [], "label": []},
...
...
{"bbox": [(y1, x1, y2, x2)], "score": [conf], "label": [cls]}]

You have to change the name of trained parameter file to model you used. For example yolov5/runs/train/<exp_no>/weights/last.pt to yolov5/runs/train/<exp_no>/weights/yolov5l.pt. Inference command:

python detect.py --source ../svhn/images/test --weights runs/train/<exp_no>/weights/yolov5l.pt --device 0 --json <output_json_file>

Increase object confidence threshold to 0.5 and infer with augmentation.

python detect.py --source ../svhn/images/test --weights runs/train/<exp_no>/weights/yolov5l.pt --device 0 --augment --conf-thres 0.5 --json <output_json_file>

Reference

• ultralytics, yolov5, viewed 25 Nov 2020, https://github.com/ultralytics/yolov5
• Bartzi, svhn_dataextract_tojson.py, viewed 25 Nov 2020, https://github.com/Bartzi/stn-ocr/blob/master/datasets/svhn/svhn_dataextract_tojson.py