I made one simplest object detection task and a model to study object detection.
- TASK
The task is to predict the bounding box of a person. Each image contains only one person. An image example is the following.
I made the dataset from MSCOCO dataset. I chose the images containing only one person, and then resized them into (32,32). I got 1k images.
- MODEL
The SimpleObjectDetector class is made as follows:
class ConvLayer(nn.Module):
def __init__(self,in_channel,out_channel):
super().__init__()
self.conv = nn.Conv2d(in_channel,out_channel,kernel_size=3,stride=1,padding=1)
self.act = nn.ReLU()
self.pool = nn.MaxPool2d(2,2)
def forward(self,x):
x = self.pool(self.act(self.conv(x)))
return x
class SimpleObjectDetector(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = ConvLayer(3,8)
self.conv2 = ConvLayer(8,16)
self.l1 = nn.Linear(8*8*16,64)
self.l2 = nn.Linear(64,4)
self.act = nn.ReLU()
def forward(self,x):
x = self.conv1(x)
x = self.conv2(x)
x = x.view(x.size()[0],-1)
x = self.act(self.l1(x))
x = self.l2(x) * 32 + 16
return x # (x,y,w,h) is expected.
x = self.l2(x) * 32 + 16 is to make the output distribution fit the image size (32,32).
- TRAINING
I tried the loss function nn.L1Loss() and IoU loss. At 1~3 epoch, the loss stops decreasing.
- RESULT
I found the predicted bounding boxes become almost the same for all images.
- ADDITIONAL STUDY
I further studied the problem, and found the bounding box prediction becomes the same for all images even before the training! It seems the final output has very low degrees of freedom.
- QUESTION
Is there any way of making this model correctly predict bounding boxes? I want to try to predict bounding boxes without using larger models like YOLO.