Comment1

.vscode/launch.json

@@ -0,0 +1,17 @@
+{
+    // Use IntelliSense to learn about possible attributes.
+    // Hover to view descriptions of existing attributes.
+    // For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
+    "version": "0.2.0",
+    "configurations": [
+        {
+            "name": "/root/anaconda3/envs/doppelgangers/bin/python",
+            "type": "python",
+            "request": "launch",
+            "program": "/root/workspace/code/ex1/DALF_CVPR_2023/train.py",
+            "console": "integratedTerminal",
+            "justMyCode": true,
+            "args": ["--mode", "ts1", "--gpu", "4,5,6,7"]
+        }
+    ]
+}

eval/dalf_benchmark.py

@@ -19,7 +19,7 @@
 
 import distmat
 
-modules = os.path.dirname(os.path.realpath(__file__)) + '/..'         
+modules = os.path.dirname(os.path.realpath(__file__)) + '/..' # 当前文件路径/../       
 sys.path.insert(0, modules)
 
 try:
@@ -35,7 +35,7 @@ def warn(*args, **kwargs):
 experiment_name = ''
 exp_dir_target = ''
 
-
+# 如果没有文件夹就创建
 def check_dir(f):
 	if not os.path.exists(f):
 		os.makedirs(f)
@@ -65,13 +65,13 @@ def parseArg():
 
 	return args
 
-
+# 标记无效位置的关键点
 def correct_cadar_csv(csv):
 	for line in csv:
 		if line['x'] < 0 or line['y'] < 0:
 			line['valid'] = 0
 
-
+# 从csv文件中提取有效位置的关键点
 def gen_keypoints_from_csv(csv):
 	keypoints = []
 	for line in csv:
@@ -82,14 +82,16 @@ def gen_keypoints_from_csv(csv):
 
 	return keypoints	 
 
-
+# 从文件中读取目录列表			
 def get_dir_list(filename):
 	with open(filename,'r') as f:
-		dirs = [line.rstrip('\n').rstrip() for line in f if line.rstrip('\n').rstrip()]
+		dirs = [line.rstrip('\n').rstrip() for line in f if line.rstrip('\n').rstrip()] # 内容中去掉换行符的非空行作为目录列表的元素
 	return dirs or False
 
+# 计算两组特征点（ref_kps 和 tgt_kps）及其描述符（ref_descriptors 和 descriptors）之间的距离
 def save_dist_matrix(ref_kps, ref_descriptors, ref_gt, tgt_kps, descriptors, tgt_gt, out_fname):
 	#np.linalg.norm(a-b)
+	# 初始化一个矩阵用于存储距离值，默认值为-1
 	print ('saving matrix in:',  out_fname)
 	size = len(ref_gt)
 	dist_mat = np.full((size,size),-1.0,dtype = np.float32)
@@ -99,36 +101,39 @@ def save_dist_matrix(ref_kps, ref_descriptors, ref_gt, tgt_kps, descriptors, tgt
 	matching_sum = 0
 
 	begin = time.time()
-
+	# 遍历两组特征点及其描述符
 	for m in range(len(ref_kps)):
 		i = ref_kps[m].class_id
+		# 检查特征点是否有效
 		if ref_gt[i]['valid'] and tgt_gt[i]['valid']:
 			valid_m+=1
 		for n in range(len(tgt_kps)):
 			j = tgt_kps[n].class_id
+			# 检查特征点是否有效
 			if ref_gt[i]['valid'] and tgt_gt[i]['valid'] and tgt_gt[j]['valid']:
+				# 计算描述符之间的距离，并存储在 dist_mat 中
 				dist_mat[i,j] = np.linalg.norm(ref_descriptors[m]-descriptors[n]) #distance.euclidean(ref_d,tgt_d) #np.linalg.norm(ref_d-tgt_d)
 
 	print('Time to match NRLFeat: %.3f'%(time.time() - begin))
-
+	# 找到每行最小值所在的列索引
 	mins = np.argmin(np.where(dist_mat >= 0, dist_mat, 65000), axis=1)
+	# 统计匹配数 TODO 有逻辑问题
 	for i,j in enumerate(mins):
 		if i==j and ref_gt[i]['valid'] and tgt_gt[i]['valid']:
 			matches+=1
 
 	print ('--- MATCHES --- %d/%d'%(matches,valid_m))
-
+	# 将距离矩阵写入文件
 	with open(out_fname, 'w') as f:
-
 		f.write('%d %d\n'%(size,size))
-
 		for i in range(dist_mat.shape[0]):
 			for j in range(dist_mat.shape[1]):
 				f.write('%.8f '%(dist_mat[i,j]))
 
-
+# TODO 比较两组关键点（kp 和 kp_gt）之间的对应关系
 def get_gt_idx(kp, kp_gt):
 	kp_dict = {}
+	# 创建字典 {k.pt:idx}
 	for idx, k in enumerate(kp):
 		kp_dict['%.2f,%.2f'%(k.pt[0],k.pt[1])] = idx
 
@@ -140,6 +145,7 @@ def get_gt_idx(kp, kp_gt):
 
 	return gt_idx
 
+# TODO 它遍历指定目录中的图像数据集，并针对指定的数据集进行DALF特征提取和匹配
 def run_benchmark(args):
 
 	dev = torch.device('cuda' if torch.cuda.is_available else 'cpu')

eval/distmat.py

@@ -1,6 +1,7 @@
 import numpy as np
 import cv2 
 
+# 计算两组描述子之间的欧氏距离矩阵，并将结果保存到文件中
 def save(desc_ref, desc_tgt, filename):
 	desc_ref = np.array(desc_ref)
 	desc_tgt = np.array(desc_tgt)
@@ -31,6 +32,7 @@ def save(desc_ref, desc_tgt, filename):
 
 		f.write('\n')
 
+# 计算两组描述子之间的欧氏距离，然后将这些距离值保存到文件中。
 def save_cvnorm(desc_ref, desc_tgt, filename):
 	desc_ref = np.array(desc_ref)
 	desc_tgt = np.array(desc_tgt)
@@ -49,6 +51,7 @@ def save_cvnorm(desc_ref, desc_tgt, filename):
 
 		f.write('\n')	
 
+# 从一个 CSV 文件中加载关键点信息并创建 OpenCV 的关键点对象列表。
 def load_cv_kps(csv):
 	keypoints = []
 	for line in csv:
@@ -57,6 +60,7 @@ def load_cv_kps(csv):
 
 	return keypoints
 
+# 保存desc到本地文件
 def save_desc(filename, desc):
 	m, n = desc.shape
 	with open(filename, 'w') as f:

eval/eval_nonrigid.sh

@@ -1,9 +1,9 @@
 #Please put here the path of images and TPS files from nonrigid benchmark
-PATH_IMGS='/srv/storage/datasets/nonrigiddataset/eval_bench/All_PNG'
-PATH_TPS='/homeLocal/guipotje/sshfs/datasets/gt_tps'
+PATH_IMGS='/root/workspace/code/ex1/DALF_CVPR_2023/dataset/eval/image/Kinect1/Bag1'
+PATH_TPS='/root/workspace/code/ex1/DALF_CVPR_2023/dataset/eval/tps/Kinect1/Bag1'
 
 #Set working dir to save results. Please change
-working_dir='/tmp/nonrigid_eval'
+working_dir='/root/workspace/code/ex1/DALF_CVPR_2023/result'
 mkdir -p $working_dir
 
 #############################################################################

eval/mean.txt

@@ -0,0 +1,6 @@
+0.10368644825861915 
+0.1264971258431803 
+0.46426299884445216 
+0.3475967964285022 
+0.24007131081249222 
+0.10698793112228643 

eval/new1.sh

@@ -0,0 +1,42 @@
+#Please put here the path of images and TPS files from nonrigid benchmark
+PATH_IMGS='/root/workspace/data/DALF/eval/image'
+PATH_TPS='/root/workspace/data/DALF/eval/tps'
+
+#Set working dir to save results. Please change
+working_dir='/root/workspace/code/ex1/DALF_CVPR_2023/result'
+mkdir -p $working_dir
+
+#############################################################################
+
+#Scripts Path
+extract_gt_path='./extract_gt.py'
+benchmark_path='./dalf_benchmark.py'
+metrics_path='./plotUnorderedPR.py'
+
+#For final eval
+ablation='model_ts1_80000_final'
+
+#Data Path
+network_path='/root/workspace/code/ex1/DALF_CVPR_2023/weights/model_ts1_80000_final.pth'
+
+#Original TPS files
+tps_dir_o=$PATH_TPS
+#Local copy of TPS files
+tps_dir=$working_dir'/gt_tps_'$ablation
+
+#Output path
+out_path=$working_dir'/out_'$ablation
+
+echo 'copying original gt_tps '$tps_dir_o' to '$tps_dir
+cp -rf $tps_dir_o $tps_dir
+python3 $extract_gt_path -i $PATH_IMGS --tps_dir $tps_dir --dir -m pgdeal --net_path $network_path
+python3 $benchmark_path -i $PATH_IMGS -o $out_path --dir --sift --tps_path $tps_dir --net_path $network_path
+#Remove old results cache
+rm *.dict
+#Show metric results
+inputdir=$out_path
+#Metric type: [MS, MMA, inliers]
+metric=MS
+# python3 $metrics_path -i $inputdir/Kinect1 -d --tps_path $tps_dir --mode erase --metric $metric
+# python3 $metrics_path -i $inputdir/SimulationICCV -d --tps_path $tps_dir --mode erase --metric $metric
+# python3 $metrics_path -i $inputdir/SimulationICCV -d --tps_path $tps_dir --mode append --metric $metric --gmean

eval/new2.sh

@@ -0,0 +1,39 @@
+#Please put here the path of images and TPS files from nonrigid benchmark
+PATH_IMGS='/root/workspace/data/DALF/eval/image'
+PATH_TPS='/root/workspace/data/DALF/eval/tps'
+
+#Set working dir to save results. Please change
+working_dir='/root/workspace/code/ex1/DALF_CVPR_2023/result'
+
+#############################################################################
+
+#Scripts Path
+extract_gt_path='./extract_gt.py'
+benchmark_path='./dalf_benchmark.py'
+metrics_path='./plotUnorderedPR.py'
+
+#For final eval
+ablation='model_ts1_80000_final'
+
+#Data Path
+network_path='/root/workspace/code/ex1/DALF_CVPR_2023/weights/model_ts1_80000_final.pth'
+
+#Original TPS files
+tps_dir_o=$PATH_TPS
+#Local copy of TPS files
+tps_dir=$working_dir'/gt_tps_'$ablation
+
+#Output path
+out_path=$working_dir'/out_'$ablation
+
+echo 'copying original gt_tps '$tps_dir_o' to '$tps_dir
+#Remove old results cache
+rm *.dict
+#Show metric results
+inputdir=$out_path
+#Metric type: [MS, MMA, inliers]
+metric=MS
+# python3 $metrics_path -i $inputdir/Kinect1 -d --tps_path $tps_dir --mode erase --metric $metric
+# python3 $metrics_path -i $inputdir/Kinect1 -d --tps_path $tps_dir --mode erase --metric $metric --gmean
+python3 $metrics_path -i $inputdir/SimulationICCV -d --tps_path $tps_dir --mode erase --metric $metric
+python3 $metrics_path -i $inputdir/SimulationICCV -d --tps_path $tps_dir --mode append --metric $metric --gmean

eval/table.txt

@@ -0,0 +1,7 @@
+& \multicolumn{1}{l}{kanagawa\_rot ($18$)}& $\underline{0.10}$  \\
+& \multicolumn{1}{l}{lascaux\_rot ($18$)}& $\underline{0.13}$  \\
+& \multicolumn{1}{l}{lascaux\_scale ($3$)}& $\underline{0.46}$  \\
+& \multicolumn{1}{l}{kanagawa\_scale ($3$)}& $\underline{0.35}$  \\
+& \multicolumn{1}{l}{chambre\_scale ($3$)}& $\underline{0.24}$  \\
+& \multicolumn{1}{l}{chambre\_rot ($18$)}& $\underline{0.11}$  \\
+

eval/table_means.txt

@@ -0,0 +1,2 @@
+& dalf
+& MS  & $0.23$