In this work, the authors demonstrate how CNNs can be used to estimate the orientation of an object between 0^o and 360^o. For this, they propose and test 3 different methods. For each method, they assume that each image comes with a single prominent object located at the center of the image. Each method is tested with a ResNet-201 pretrained on ImageNet.

Method 1

Orientation is defined by a vector \(\vec v=(cos(\theta),sin(\theta))\) and the loss is a L1 norm. During testing, the predicted vector \(\vec v=(x,y)\) is converted to an angle with a atan2 function.

Method 2

Same as Method 1 but with a cosine loss \(L(\vec v_{gt},\vec v)=1-cos(\theta)\) where theta is the angle between the predicted vector and the groundtruth vector.

Method 3

The third method uses a finite number of discreet \(N=4\) orientations. In order to reduce the discretization error, they train 3 different networks with different starting orientation angle as shown in fig.2. The softmax prediction of the 3 models are then combined with a mean-shift method and the orientation with the maximum probability is retained.

Results

They tested their method on 2 datasets : EPFL-Car and TUD-Pedestrian. In both cases, Method 3 beats Method 1 and 2 as well as previous works.

Example of results

Code

Code is available here