TextBoxes re-implementation using tensorflow.
Much more info can be found in Textbox and SSD
This project is greatly inspired by slim project
And many functions are modified based on SSD-tensorflow project
Now the pipeline is much clear and can be resued in any tf projects.
Author: Daitao Xing : [email protected] Jin Huang : [email protected]
For now, the model can detect most of boxes. But still has poor performance on small objects.
- download the sythtext data from link
- Unzip the datasets and put all folders and files under
/data/sythtext/ cd datasetsand run scriptpython data2record.py- The ICDAR2013 datasets follow the same pipeline.
You can train this model from scratch by using following command.
DATASET_DIR=./data/sythtext/ TRAIN_DIR=./logs/train CUDA_VISIBLE_DEVICES=0,1,2,3 setsid python Textbox_train.py \ --train_dir=${TRAIN_DIR} \ --dataset_dir=${DATASET_DIR} \ --save_summaries_secs=60 \ --save_interval_secs=1800 \ --weight_decay=0.0005 \ --optimizer=momentum \ --learning_rate=0.001 \ --batch_size=8 \ --num_samples=800000 \ --gpu_memory_fraction=0.95 \ --max_number_of_steps=500000 \ --use_batch=False \ --num_clones=4 \
You can download vgg checkpoint from web and put it under checkpoints folder. To train the model from checkpoint(suggested), please run the following lines
# or change into your checkpoint file name
CHECKPOINT_PATH=./checkpoints/vgg_16.ckpt
DATASET_DIR=./data/sythtext/
TRAIN_DIR=./logs/train
CUDA_VISIBLE_DEVICES=0,1,2,3 setsid python Textbox_train.py \
--train_dir=${TRAIN_DIR} \
--dataset_dir=${DATASET_DIR} \
--save_summaries_secs=60 \
--save_interval_secs=1800 \
--weight_decay=0.0005 \
--learning_rate=0.001 \
--batch_size=8 \
--num_samples=800000 \
--gpu_memory_fraction=0.42 \
--max_number_of_steps=500000 \
--use_batch=False \
--num_clones=4 \
--checkpoint_path=${CHECKPOINT_PATH} \
--checkpoint_model_scope=vgg_16 \
--ignore_missing_vars=True \
CHECKPOINT_PATH=./logs/train/logs614
EVAL_DIR=./logs/eval/logs614
DATASET_DIR=./data/ICDAR2013/test
CUDA_VISIBLE_DEVICES=4 setsid python eval.py \
--eval_dir=${EVAL_DIR} \
--dataset_dir=${DATASET_DIR} \
--checkpoint_path=${CHECKPOINT_PATH} \
--wait_for_checkpoints=True \
--batch_size=1 \
--gpu_memory_fraction=0.02 \
--use_batch=False \For further instruction on how to use slim. please check on slim project.
- Anchor boxes size is very important. Many small objects can't be matched with any anchor boxes. This is a drawback of SSD.
- BatchNorm help a lot. Batch norm will imporove accurary by 5-10% percentage.
- Tensorflow is ineffecient compared with Caffe. For now this pipiline is under optimization. The speed is about 3 sec/step on sigle gpu(K80) where batch size is 32.
- Training as long as you can. I can't reproduce the result from scratch with only training 50K times. The result above is the result training after 150K