logos.tensorflow

The Problem

We would like to detect brand logos in an image.

TODO

• Project setup
• Data Preprocessing
• Train models
• Save results
• Write readme with results
• Add training plots to readme
• Tests
• CI/CD
• API
• Deploy
• Train with distractor logo images

Setup Project

I recommended use docker and work inside container

docker build -f ./dockerfiles/gpu/Dockerfile -t logost.tensorflow .
nvidia-docker run -it -v $PWD:/code --net=host --ipc=host logost.tensorflow:latest /bin/bash

Dataset

You can find origin dataset here Also you can download my train/val/test split with tf.records here

Or you can create your own dataset

mkdir dataset
./scripts/load_data.sh
python ./logos/data_preprocessing.py
python ./logos/create_tf_records.py

Results

	name	speed (ms)	train mAP	val mAP	test images	pipeline	checkpoints
0	ssd_mobilenet_v1	30	0.9088	0.6811	load	pipeline	load
1	ssd_mobilenet_v1_focal_loss	30	0.9518	0.671	load	pipeline	load
2	faster_rcnn_resnet50	89	0.9893	0.7395	load	pipeline	load

Commands for tensorflow object detection API

Parameters

export PATH_TO_YOUR_PIPELINE_CONFIG=...
export PATH_TO_TRAIN_DIR=...
export PATH_TO_EVAL_DIR=...
export TRAIN_PATH=...
export EXPORT_DIR=...

Train/Eval/Export model

python object_detection/train.py \
    --logtostderr \
    --pipeline_config_path=${PATH_TO_YOUR_PIPELINE_CONFIG} \
    --train_dir={PATH_TO_TRAIN_DIR}

python object_detection/eval.py \
    --logtostderr \
    --pipeline_config_path=${PATH_TO_YOUR_PIPELINE_CONFIG} \
    --checkpoint_dir=${PATH_TO_TRAIN_DIR} \
    --eval_dir=${PATH_TO_TRAIN_DIR}

python object_detection/export_inference_graph.py \
    --input_type image_tensor \
    --pipeline_config_path=${PATH_TO_YOUR_PIPELINE_CONFIG} \
    --trained_checkpoint_prefix=${TRAIN_PATH} \
    --output_directory=${EXPORT_DIR}

References

• Speed/accuracy trade-offs for modern convolutional object detectors
• Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
• Focal Loss for Dense Object Detection
• A systematic study of the class imbalance problem in convolutional neural networks
• SSD: Single Shot MultiBox Detector
• MobileNetV2: Inverted Residuals and Linear Bottlenecks
• Faster R-CNN Explained
• Understanding SSD MultiBox — Real-Time Object Detection In Deep Learning