jeffkang-94/pytorch-ssd300

Pytorch Implementation for SSD300

Pytorch Implementation of SSD300

We redesign and fix the bug in original implementation which considers pytorch 0.4.

This code supports pytorch 1.0 > in python 3.6.

plz, refer to detail information on paper

Objective

To build a model that can detect and localize specific objects in images.

This repository addresses Single Shot Multibox Detector (SSD), a popular, powerful, and especially nimble network for this task. The authors' original implementation can be found here.

Usage

Quick overview of entire procedure.
We elaborate on the procedure with the following sections.

Overall Training

cd asset
bash download_voc.sh
cd ..
python create_data_list.py
python train.py

Overall Test

cd asset
bash download.sh
cd ..
python detect.py or python eval.py

Dataset

We use VOC2007 and VOC2012 dataset to train SSD300.
You can download those datasets using below command.

cd asset
bash download_voc.sh

VOCdevkit
-| VOC2007
   -| Annotations
   -| ImageSets
   -| JPEGImages
   -| SegmentationClass
   -| SegmentationObject
-| VOC2012
   -| Annotations
   -| ImageSets
   -| JPEGImages
   -| SegmentationClass
   -| SegmentationObject

Create Data List

Before you train the model, you need to preprocess the data.
Specify the data root in create_data_list.py.

from utils import create_data_lists

if __name__ == '__main__':
    create_data_lists(voc07_path='[VOC2007 Datapath]', # specify your data root
                      voc12_path='[VOC2012 Datapath]',
                      output_folder='./')

python create_data_list.py

then, TRAIN_images.json TEST_images.json and TRAIN_objects.json TEST_objects.json files are generated.

Training

If the json files were successfully generated, you can now train the SSD300 model.

python train.py 

We use SGD optimizer with momentum=0.9 and adopt lr decay at 80000,100000 iteration. grad_clip is useful if you afford to use a large batch size (e.g., more than 32). We train the model batch_size=8 in single TITAN RTX without grad_clip.

Refer to the training setting as below:

# Learning parameters
checkpoint = None  # path to model checkpoint, None if none
batch_size = 8  # batch size
iterations = 120000  # number of iterations to train
workers = 4  # number of workers for loading data in the DataLoader
print_freq = 200  # print training status every __ batches
lr = 1e-3  # learning rate
decay_lr_at = [80000, 100000]  # decay learning rate after these many iterations
decay_lr_to = 0.1  # decay learning rate to this fraction of the existing learning rate
momentum = 0.9  # momentum
weight_decay = 5e-4  # weight decay
grad_clip = None  # clip if gradients are exploding, which may happen at larger batch sizes (sometimes at 32) - you will recognize it by a sorting error in the MuliBox loss calculation

Test

We provide a pre-trained model with link.
You can download it using above link or using shell file download.sh

cd asset
bash download.sh

You can detect objects based on single image using detect.py.
Given the path of single image, you can process the object detection using pre-trained model and save the result.

if __name__ == '__main__':
    img_path = '[Path of single image]' # e.g., /mnt2/datasets/VOCdevkit/VOC2007/JPEGImages/000131.jpg
    original_image = Image.open(img_path, mode='r')
    original_image = original_image.convert('RGB')
    annotated_image = detect(original_image, min_score=0.2, max_overlap=0.5, top_k=200)
    annotated_image.save('[Name of result image]') # e.g., ./result.jpg

Evaluation

For evaluation, SSD model use mAP(mean Average Precison).
Detail for how calculate the mAP is provided in calculate_mAP in utils.py
It takes a few minutes (definetely depends on your environment).
We obtain mAP 70.2%.

python eval.py

Model	data	mAP	aero	bike	bird	boat	bottle	bus	car	chair	cow	table	dog	horse	mbike	person	plant	sheep	sofa	train	tv
SSD300	VOC07+12	74.3	75.5	80.2	72.3	66.3	47.6	83.0	84.2	86.1	54.7	78.3	73.9	84.5	85.3	82.6	76.2	48.6	73.9	76.0	83.4
Trial_1	VOC07+12	70.2	70.1	80.4	64.8	61.7	39.6	81.1	80.1	79.2	51.0	75.6	74.3	75.2	81.4	79.1	73.4	41.5	71.7	73.8	82.9
Trial_2	VOC07+12

Demo images

pytorch-ssd300's Issues

utils.py image normalization part annotation reason

pytorch-SSD300/utils.py

Line 611 in 6cb6b3c

#new_image = FT.normalize(new_image, mean=mean, std=std)

안녕하세요 ssd code review하면서 Jeffkang님의 코드로 많은 도움 받았습니다. 감사합니다.
궁금한 점이 있는데, 이 부분 주석처리한 이유가 있을까요?
(imagenet std 와 mean값을 이용해서 normalize를 해야된다고 생각하고 있어서요)