wave

Author: curme-miller

README.md

@@ -8,7 +8,7 @@
 | Component | Status |
 |----------------------------------|----------------|
 | 🧠 Model Definition | ✅ Available |
-| 🏋️‍♀️ Training Code | ✅ Available  |
+| 🏋️‍♀️ Training Code | ⏳ Coming Soon |
 | 🧪 Inference Code | ✅ Available |
 | 🎯 Model Weights | ⏳ Coming Soon |
 | 📊 Dataset | ⏳ Coming Soon |
@@ -20,8 +20,9 @@
 - **Status**: Under Review at [AAAI 2026]
 
 ---
+
 ## Model Weights and Datasets
-**Model Weights** and **Datasets** will be made public after the paper is accepted. (*Coming Soon!!!*)
+**Training Code**, **Model Weights** and **Datasets** will be made public after the paper is accepted. (*Coming Soon!!!*)
 
 ## Abstract
 Accurately localizing and segmenting occluded objects from faint light patterns beyond the field of view is highly challenging due to multiple scattering and medium-induced perturbations. Most existing methods, based on real-valued modeling or local convolutional operations, are inadequate for capturing the underlying physics of coherent light propagation. Moreover, under low signal-to-noise conditions, these methods often converge to non-physical solutions, severely compromising the stability and reliability of the observation. To address these challenges, we propose a novel physics-driven Wavefront Propagating Compensation Network (WavePCNet) to simulates wavefront propagation and enhance the perception of occluded objects. This WavePCNet integrates the Tri-Phase Wavefront Complex-Propagation Reprojection (TriWCP) to incorporate complex amplitude transfer operators to precisely constrain coherent propagation behavior, along with a momentum memory mechanism to effectively suppress the accumulation of perturbations. Additionally, a High-frequency Cross-layer Compensation Enhancement is introduced to construct frequency-selective pathways with multi-scale receptive fields and dynamically models structural consistency across layers, further boosting the model’s robustness and interpretability under complex environmental conditions. Extensive experiments conducted on four physically collected datasets demonstrate that WavePCNet consistently outperforms state-of-the-art methods across both accuracy and robustness. All data and code will be publicly released to support and encourage continued research in the obscured object detection domain.
@@ -37,5 +38,7 @@ Accurately localizing and segmenting occluded objects from faint light patterns
 ## Result
 ![](./img/result.png)
 
+---
 
+## Test
 

base/config.py

@@ -0,0 +1,87 @@
+import sys 
+import os
+import argparse 
+import importlib
+from .util import *
+
+def base_config(net_name):
+ parser = argparse.ArgumentParser()
+ model_dafault_config = importlib.import_module('methods.{}'.format(net_name)).custom_config
+
+ parser.add_argument('model_name', default=net_name, help='Model name')
+ parser.add_argument('--backbone', default='resnet50', help='Set the backbone of the model')
+ parser.add_argument('--show_param', action='store_true') # show the number of parameter
+ 
+ # Training schedule
+ parser.add_argument('--sub', default='base', help='Job name')
+ parser.add_argument('--clip_gradient', default=0, type=float, help='Max gradient')
+ parser.add_argument('--weight_decay', default=0.0005, type=float)
+ parser.add_argument('--data_aug', action='store_false', help='Data augmentation, only random crop')
+ parser.add_argument('--multi', action='store_false', help='Multi-scale training')
+ parser.add_argument('--gpus', default='0', type=str, help='Set the gpu devices')
+ parser.add_argument('--strategy', default='adam_base', help='Training strategy, see base/strategy.py')
+ parser.add_argument('--batch', default=10, type=int, help='Batch Size for Testing')
+
+ parser.add_argument('--training_stage', default=2, type=int, help='Training stage: 1=physical only, 2=physical+network, 3=freeze physical')
+ # Data setting
+ parser.add_argument('--size', default=384, type=int, help='Input size')
+ parser.add_argument('--trset', default='DUTS-TR', help='Set the traing set')
+ parser.add_argument('--vals', default='all', help='Set the testing sets') 
+ parser.add_argument('--data_path', default='/home/v2-4080s/ouyang/OOD/kong/data/亮度/', help='Dataset path')
+ parser.add_argument('--save_path', default='./result/our/HKU/meanetA更改/', help='Save path')
+ parser.add_argument('--weight_path', default='./weight/HKU/meanetA更改/', help='Weight path')
+ 
+
+ # Testing
+ parser.add_argument('--resume', action='store_true')
+ parser.add_argument('--weight', default='', help='Loading weight file')
+ parser.add_argument('--save', action='store_true', help='Whether save result')
+ parser.add_argument('--test_batch', default=6, type=int, help='Batch Size for Testing')
+ parser.add_argument('--debug', action='store_true') # Test model before training
+ 
+ # Use for SALOD dataset
+ parser.add_argument('--train_split', default=10000, type=int, help='Use for SALOD dataset')
+ 
+ # Construct loss by loss_factory. More details in base/loss.py.
+ parser.add_argument('--loss', default='bce,iou', type=str, help='Losses for networks')
+ parser.add_argument('--lw', default='1,1', type=str, help='Weights for losses')
+ 
+ # Customized arguments
+ ### Base arguments with customized values
+ parser.set_defaults(**model_dafault_config['base'])
+ 
+ ### Customized arguments
+ for k, v in model_dafault_config['customized'].items():
+ v['dest'] = k[2:]
+ parser.add_argument(k, **v)
+ 
+ params = parser.parse_args()
+ config = vars(params)
+
+ if config['trset'] == 'SALOD':
+ config['vals'] = ['SALOD']
+ elif config['trset'] == 'simple':
+ config['vals'] = ['tough', 'normal']
+ elif config['trset'] == 'DUTS-TR':
+ if config['vals'] == 'all':
+ config['vals'] = ['PASCAL-S', 'ECSSD', 'HKU-IS', 'DUTS-TE', 'DUT-OMRON']
+ else:
+ config['vals'] = config['vals'].split(',')
+
+ elif config['trset'] == 'COD-TR':
+ if config['vals'] == 'all':
+ config['vals'] = ['COD-TE', 'NC4K', 'CAMO-TE']
+ else:
+ config['vals'] = config['vals'].split(',')
+ else:
+ config['vals'] = config['vals'].split(',')
+ 
+ save_path = os.path.join(config['save_path'], config['model_name'], config['backbone'], config['sub'])
+ check_path(save_path)
+ config['save_path'] = save_path
+ 
+ weight_path = os.path.join(config['weight_path'], config['model_name'], config['backbone'], config['sub'])
+ check_path(weight_path)
+ config['weight_path'] = weight_path
+ 
+ return config

base/data.py

@@ -0,0 +1,138 @@
+import os
+import torch
+import random
+import numpy as np
+from PIL import Image
+import torch.utils.data as data
+import torchvision.transforms as transforms
+
+mean = np.array([0.485, 0.456, 0.406]).reshape([1, 1, 3])
+std = np.array([0.229, 0.224, 0.225]).reshape([1, 1, 3])
+
+def get_image_list(name, config, phase):
+ images = []
+ gts = []
+
+ if name == 'all':
+ print("Skipping 'all' dataset as it does not have a corresponding directory.")
+ return images, gts
+
+ if name in ('simple', 'tough', 'normal'):
+ train_split = 10000
+ 
+ print('Objectness shifting experiment.')
+ # Objectness
+ list_file = 'clean_list.txt'
+ f = open(os.path.join(config['data_path'], 'SALOD/{}'.format(list_file)), 'r')
+ if name == 'simple':
+ img_list = f.readlines()[-train_split:]
+ elif name == 'normal':
+ img_list = f.readlines()[train_split:-train_split]
+ else:
+ img_list = f.readlines()[:train_split]
+ 
+ for i in range(len(img_list)):
+ img_list[i] = img_list[i].split(' ')[0]
+ 
+ images = [os.path.join(config['data_path'], 'SALOD/images', line.strip() + '.jpg') for line in img_list]
+ gts = [os.path.join(config['data_path'], 'SALOD/mask', line.strip() + '.png') for line in img_list]
+ 
+ # Benchmark
+ elif name == 'SALOD':
+ f = open(os.path.join(config['data_path'], 'SALOD/{}.txt'.format(phase)), 'r')
+ img_list = f.readlines()
+ 
+ images = [os.path.join(config['data_path'], name, 'images', line.strip() + '.jpg') for line in img_list]
+ gts = [os.path.join(config['data_path'], name, 'mask', line.strip() + '.png') for line in img_list]
+
+ elif phase == 'test' and os.path.isabs(name):
+ image_root = os.path.join(name, 'images')
+ gt_root = os.path.join(name, 'segmentations')
+ 
+ images = sorted([os.path.join(image_root, f) for f in os.listdir(image_root) if f.endswith('.jpg') or f.endswith('.png')])
+ gts = sorted([os.path.join(gt_root, f) for f in os.listdir(gt_root) if f.endswith('.png')]) 
+
+ else:
+ image_root = os.path.join(config['data_path'], name, 'images')
+ print("Name:", name)
+ print("Image root:", image_root)
+ gt_root = os.path.join(config['data_path'], name, 'segmentations')
+ 
+ images = sorted([os.path.join(image_root, f) for f in os.listdir(image_root) if f.endswith('.jpg') or f.endswith('.png')])
+ gts = sorted([os.path.join(gt_root, f) for f in os.listdir(gt_root) if f.endswith('.png')])
+ 
+ return images, gts
+
+def get_loader(config):
+ dataset = Train_Dataset(config['trset'], config)
+ data_loader = data.DataLoader(dataset=dataset,
+ batch_size=config['batch'],
+ shuffle=True,
+ num_workers=12,
+ pin_memory=True,
+ drop_last=True)
+ return data_loader
+
+
+
+class Train_Dataset(data.Dataset):
+ def __init__(self, name, config):
+ self.config = config
+ self.images, self.gts = get_image_list(name, config, 'train')
+ self.size = len(self.images)
+
+ def __getitem__(self, index):
+ image = Image.open(self.images[index]).convert('RGB')
+ gt = Image.open(self.gts[index]).convert('L')
+ 
+ img_size = self.config['size']
+ image = image.resize((img_size, img_size))
+ gt = gt.resize((img_size, img_size))
+ 
+ image = np.array(image).astype(np.float32)
+ gt = np.array(gt)
+ 
+ image = ((image / 255.) - mean) / std
+ image = image.transpose((2, 0, 1))
+ gt = np.expand_dims((gt > 128).astype(np.float32), axis=0)
+
+ image = torch.from_numpy(image).float()
+ gt = torch.from_numpy(gt).float()
+ return image, gt
+
+ def __len__(self):
+ return self.size
+
+class Test_Dataset:
+ def __init__(self, name, config=None):
+ self.config = config
+ self.images, self.gts = get_image_list(name, config, 'test')
+ self.size = len(self.images)
+
+ def load_data(self, index):
+ image = Image.open(self.images[index]).convert('RGB')
+ image = image.resize((self.config['size'], self.config['size']))
+ image = np.array(image).astype(np.float32)
+ gt = np.array(Image.open(self.gts[index]).convert('L'))
+ gt = Image.open(self.gts[index]).convert('L')
+ gt = gt.resize((self.config['size'], self.config['size']))
+ gt = np.array(gt).astype(np.float32)
+ name = self.images[index].split('/')[-1].split('.')[0]
+ 
+ 
+ image = ((image / 255.) - mean) / std
+ image = image.transpose((2, 0, 1))
+ image = torch.tensor(np.expand_dims(image, 0)).float()
+ gt = (gt > 128).astype(np.float32)
+ return image, gt, name
+
+def test_data():
+ config = {'orig_size': True, 'size': 288, 'data_path': '../dataset'}
+ dataset = 'SOD'
+ data_loader = Test_Dataset(dataset, config)
+ imgs, gts, names = data_loader.load_all_data()
+ print(imgs.shape, gts.shape, len(names))
+ 
+
+if __name__ == "__main__":
+ test_data()