我正在尝试在PascalVoc预训练数据集上运行CNTK对象检测示例。我在fastrcnn中运行所有必需的脚本，并获取数据集中定义的测试数据的可视输出。现在我想在我自己的图像上测试网络，我该怎么做？
答案 0 :(得分：0)
对于快速R-CNN，您需要一个库，为您的测试图像生成候选ROI（感兴趣的区域），例如： selective search。
如果您想评估一批图片，可以按照tutorial中的说明生成测试映射文件和ROI坐标（请参阅相应的test.txt和test.rois.txt proc子文件夹）。如果要评估单个图像，则需要将图像和候选ROI坐标作为cntk eval的输入传递，类似于this example：
# compute model output
arguments = {loaded_model.arguments[0]: [hwc_format]}
output = loaded_model.eval(arguments)
答案 1 :(得分：0)
对于FastRCNN，您需要首先通过选择性搜索算法运行自定义图像以生成ROI（感兴趣的区域），然后将其提供给您的模型，如下所示：
using namespace std;
int m, n;
string walls(int a){
// west north east south
if(a == 1){
return "1000";
}
else if(a == 2){
return "0100";
}
else if(a == 3){
return "1100";
}
else if(a == 4){
return "0010";
}
else if(a == 5){
return "1010";
}
else if(a == 6){
return "0110";
}
else if(a == 7){
return "1110";
}
else if(a == 8){
return "0001";
}
else if(a == 9){
return "1001";
}
else if(a == 10){
return "0101";
}
else if(a == 11){
return "1101";
}
else if(a == 12){
return "0011";
}
else if(a == 13){
return "1011";
}
else if(a == 14){
return "0111";
}
else{
return "1111";
}
}
//checking if a square isn't just four walls
bool isOpen(string s){
if(s.find('0') != string::npos){
return true;
}
return false;
}
bool isRight(string x, string y){
//if there's a east well open for x, y must have a west wall open
if((x[2] == '0')&&(y[0] == '0')){
return true;
}
return false;
}
bool isDown(string x, string y){
//if there's a south wall open for x, y must have a north wall open
if((x[3] == '0')&&(y[1] == '0')){
return true;
}
return false;
}
//size of the room that you're finding the path in
int roomSize = 0;
string seenCoords;
int findPath(string modules[n][m], int i, int j){
if((i < n)&&(j < m)){
if(isOpen(modules[i][j])){
//use dashes to keep track of ij coords
//13-15-23-
//13 15 and 23 are considered ij coords
//checking whether square i,j has been seen before
if(seenCoords.find(to_string(i) + to_string(j) + "-") == string::npos){
//add the coords to the ones you've already seen
seenCoords += to_string(i) + to_string(j) + "-";
roomSize++;
if(isRight(modules[i][j], modules[i+1][j])){
roomSize += findPath(modules, i+1, j);
}
if(isDown(modules[i][j], modules[i][j+1])){
roomSize += findPath(modules, i, j+1);
}
}
}
}
//last char in string is the room size
return roomSize;
}
int main(){
ifstream fin("castle.in");
ofstream fout("castle.out");
fin >> m >> n;
string modules[n][m];
for(int i = 0; i < n; i++){
for(int j = 0; j < m; j++){
int a;
fin >> a;
modules[i][j] = walls(a);
}
}
int roomCount, largestRoom = 0;
for(int i = 0; i < n; i++){
for(int j = 0; j < m; j++){
//if it's not just four walls
if(isOpen(modules[i][j])){
//if you haven't seen the square i, j before
if(seenCoords.find(to_string(i) + to_string(j) + "-") == string::npos){
//recurse on the square and find all the paths off it
roomSize = findPath(modules, i, j);
roomCount++;
if(roomSize > largestRoom){
largestRoom = seenCoords[seenCoords.length()-1];
}
roomSize = 0;
}
}
else{
//if it's not open it's just made of four walls
roomCount++;
if(largestRoom < 1){
largestRoom = 1;
}
}
}
}
}
您可以在此处找到更多详细信息：https://github.com/Microsoft/CNTK/tree/release/latest/Examples/Image/Detection/FastRCNN
无论如何，由于使用选择性搜索（这是一个真正的瓶颈），FastRCNN不是最有效的方法。如果你想提高性能，你可以尝试FasterRCNN，因为它摆脱了SS算法，并用Region Proposal Network取代它，它可以做得更好。 如果您有兴趣，可以在GitHub上查看我的回购：https://github.com/karolzak/CNTK-Hotel-pictures-classificator