Developers: Kaney Nguyen & Martin Navarrete
Figure 1 - Camouflaged with a pyramid
Figure 2 - Ground Truth
- A camouflaged object refers to an item or entity that is intentionally designed, colored, or disguised to blend in with its surroundings, in order to conceal its presence or to make it difficult to be detected by visual observation. Camouflaged objects are often used in military, hunting, or wildlife photography contexts, where stealth or concealment is crucial. The object may be concealed using natural or artificial materials, such as foliage, paint, netting, or other materials, that are carefully chosen and arranged to match the texture, color, and pattern of the environment. The purpose of camouflaging an object is to make it blend in so well that it becomes nearly invisible, making it difficult to distinguish from the surrounding objects or landscape.
- This software is intergrating with PFNET. For more info about the DNN, see paper
- The goal of this project is to build a software that is easy to use. Many camouflaged object detection softwares lacks of end-user experience, and very difficult to use for average people.
demotest-p0q1pngh_8u3Cyz63.1.mp4
- Medical
- Millitary
- Environmental
- MacOS(M1/M2 Chipset)
- Window
- Linux
PFNet_Ease
├── data
│ | |
│ ├── Train
| | ├── Project_Name
│ │ | ├── image
│ │ | ├── gts
│ ├── Test
| | ├── Project_Name
│ │ | ├── image
│ │ | ├── gts
.
.
.
If you have trouble to install the pytorch use this code below:
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 -f https://download.pytorch.org/whl/torch_stable.html
# clone the git hub
git clone PFNet_Ease
# go to code folder
cd PFNet_Ease
# apt install required packages
pip install -r requirementspython train.py --help
-h, --help show this help message and exit
--img_size IMG_SIZE The size of the input image. Adjust the
train scale by specifying a value, for
example: 416, 704, or 1024. Default:
704
--batch_size BATCH_SIZE The number of samples in each batch.
Adjust as needed. Default: 32
--epochs EPOCHS The number of epochs to train for.
Adjust as needed. Default: 100
--last_epoch LAST_EPOCH The index of the last epoch completed.
Default: 0
--lr LR The learning rate for training.
Default: 0.001
--optimizer OPTIMIZER The optimizer to use for training. For
example: Adam. Default: SGD
--weight_decay WEIGHT_DECAY The weight decay for regularization
during training. Default: 0.0005
--snap_shot SNAP_SHOT Resume Training or Transfer Training
Default:
--lr_decay LR_DECAY The learning rate decay factor.
--momentum MOMENTUM The momentum for optimizer during
training. Default: 0.9
--poly_train POLY_TRAIN Set to True for polynomial decay
learning rate during training. Default:
True
--save_point SAVE_POINT Epochs at which to save the model
weights. Enter as a list, for example:
[1,10,20,30,40,50]. Default: [1, 10,
20, 30, 40, 50]
--train_path TRAIN_PATH The path to the training data. Default:
train/custom
--dataset_path DATASET_PATH The path to the root of the dataset.
Default: ./data/
--exp_name EXP_NAME Set experiment name. Default: test1
--project_name PROJECT_NAME Set project name. Default: project1
--ckpt_path CKPT_PATH The location where you want to save
your training result. Default: ./ckpt
--test_path TEST_PATH The path to the test data. Default:
test/custom
--result_path RESULT_PATH The path to the results. Default:
./results/
--load_weight LOAD_WEIGHT The path to the pre-trained weight
file. Enter in the format: Path +
weight_name.pth. Default: ./best.pth
--frame_scale FRAME_SCALE The percentage by which to upscale or
downscale the camera capture. Default:
100
--load_video LOAD_VIDEO The path to the video file to load.
Enter in the format: "./monkey.mp4".
Default: videoname.mp4
--select_camera SELECT_CAMERA The index of the camera to use. Enter
as an integer value from 0 to 4.
Default: 0
--display_accuracy DISPLAY_ACCURACY Display TP, TN, FP, FN, Accuracy. Note:
Required Proper Formatting to work.
Default: False
--display_area DISPLAY_AREA Display area accuracy. Note: Required
Proper Formatting to work. Default:
False
--device DEVICE The index of the GPU. For example: 0,
1, 2, or 3. Default: 0
--save_video SAVE_VIDEO Save result video Default: False
--save_results SAVE_RESULTS Save infer result Default: True
--num_workers NUM_WORKERS The number of worker threads to use.
Default: 16
Data preparation:
Prepare image with camouflaged and ground truth of the camouflage image:
Create two folders:
image and gts as the folder structure aboveExample of camouflaged dataset:
Download Camouflaged Example
Example:
python train.py --batch_size 8 --device 0 --exp_name "Pyramid" --epoch 100 --dataset_path ./data --train_path Pyramid --img_size 325
Training Log Example:
warnings.warn(_create_warning_msg(
Namespace(img_size=704, batch_size=8, epochs=100, last_epoch=0, lr=0.001, optimizer='SGD', weight_decay=0.0005, snap_shot='', lr_decay=0.9, momentum=0.9, poly_train=True, save_point=[1, 10, 20, 30, 40, 50], device=0, train_path='Train/custom_camo', dataset_path='./data', exp_name='custom_camo', ckpt_path='./ckpt', test_path='./test/Train_Mix', load_weight='./best.pth', frame_scale=100, load_video='videoname.mp4', select_camera=0, multi_gpu=True, num_workers=16)
{'epoch_num': 100, 'train_batch_size': 8, 'last_epoch': 0, 'lr': 0.001, 'lr_decay': 0.9, 'weight_decay': 0.0005, 'momentum': 0.9, 'snapshot': '', 'scale': 704, 'save_point': [1, 10, 20, 30, 40, 50], 'poly_train': True, 'optimizer': 'SGD'}
custom_camo
From ./backbone/resnet/resnet50-19c8e357.pth Load resnet50 Weights Succeed!
SGD
Using 1 GPU(s) to Train.
0%| | 0/134 [00:00<?, ?it/s]/usr/local/lib/python3.9/dist-packages/torch/nn/_reduction.py:42: UserWarning: size_average and reduce args will be deprecated, please use reduction='mean' instead.
warnings.warn(warning.format(ret))
[ 1], [ 134], [0.000991], [8.68393], [1.17123], [1.08682], [1.07181], [1.0705
[ 2], [ 268], [0.000982], [6.92324], [0.90088], [0.84093], [0.82352], [0.8836
[ 3], [ 402], [0.000973], [5.81829], [0.79145], [0.73654], [0.71588], [0.7146
[ 4], [ 536], [0.000964], [5.38071], [0.73253], [0.68674], [0.66433], [0.6582
[ 5], [ 670], [0.000955], [5.15818], [0.69671], [0.66556], [0.63779], [0.6300
[ 6], [ 804], [0.000946], [5.12762], [0.71141], [0.66046], [0.63120], [0.6233
[ 7], [ 938], [0.000937], [5.01938], [0.68252], [0.65180], [0.61902], [0.6117
[ 8], [ 1072], [0.000928], [4.96626], [0.67320], [0.64332], [0.61346], [0.6057
[ 9], [ 1206], [0.000919], [4.92973], [0.66921], [0.63748], [0.60827], [0.6016
[ 10], [ 1340], [0.000910], [4.92222], [0.66695], [0.63663], [0.60796], [0.6006
[ 11], [ 1474], [0.000900], [4.89363], [0.66599], [0.63229], [0.60392], [0.5968
[ 12], [ 1608], [0.000891], [4.86323], [0.66102], [0.62732], [0.60032], [0.5935
[ 13], [ 1742], [0.000882], [4.84930], [0.65632], [0.62620], [0.59896], [0.5922
[ 14], [ 1876], [0.000873], [4.85351], [0.65462], [0.62585], [0.59979], [0.5933
[ 15], [ 2010], [0.000864], [4.81631], [0.65104], [0.62109], [0.59520], [0.5884
[ 16], [ 2144], [0.000855], [4.81192], [0.65121], [0.61958], [0.59446], [0.5880
[ 17], [ 2278], [0.000846], [4.79083], [0.64637], [0.61654], [0.59249], [0.5857
[ 18], [ 2412], [0.000836], [4.78060], [0.64511], [0.61487], [0.59132], [0.5845To Resume or transfer the learning. Simply type in the epoch number.
Example: Resume at epoch 1 aka 1.pth
python train --snapshot 1
Example:
python infer.py --load_weight ./best-k-1.pth --test_path New/test/k-1 --gt_object 1
NOTE: You must have gts folder in test or it would crash. Change gt_object 2 if there are 2 objects in the ground truth.
Test log example:
test1's average Time Is: 0.053 s
test1's average Time Is: 19.0 fps
True positive: 65.0%
True negative: 0.0%
False positive: 35.0%
False negative: 0.0%
Accuracy: 65.0%
Calculate Mask Area Accuracy...
The average accuracy for contour area is: 82.99%
Total Testing Time: 0:01:16
Figure 3 - Result Mask
On image:
python infer.py --load_weight ./best-k-1.pth --test_path New/test/k-1
Figure 4 - Result
python detect_video.py --load_video video_path --load_weight weight_path
python detect_camera.py --load_weight weight_path --frame_scale 100@InProceedings{Mei_2021_CVPR,
author = {Mei, Haiyang and Ji, Ge-Peng and Wei, Ziqi and Yang, Xin and Wei, Xiaopeng and Fan, Deng-Ping},
title = {Camouflaged Object Segmentation With Distraction Mining},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2021},
pages = {8772-8781}
}
@Software{PFNet_Ease,
author = {Kaney Nguyen, Martin Navarrete},
title = {PFNet_Ease},
webpage = {https://github.com/MagicKey23/PFNet_Ease/},
month = {April},
year = {2023}
}
PFNET_Ease is available under two different licenses:
- PFNET License: See LICENSE file for details.
- MIT License: Refer to License.txt
- The source code is free for research and education use only. Any commercial usage should get formal permission first.