In general, thermal images have a poor resolution, which could be improved by using learning-based traditional super-resolution methods. These methods have been largely used in the visible spectral domain. They work by downsampling and adding noise and blur to the given image. These noisy and blurred poor quality images, together with the given images (which are used as the Ground Truths), are used in the learning process.

The approach mentioned above has been mostly used to tackle the super-resolution problem, however there are few contributions where the learning process is based on the usage of a pair of images (low and a high-resolution images) obtained from different cameras. For the current challenge, a novel thermal image dataset has been created, containing images with three different resolutions (low, mid, high) obtained with three different thermal cameras. The challenge consists in creating a solution capable of generating super-resolution images in x2, x3 and x4 scales from the different resolutions, in the case of x2 an additional evaluation will be performed by using a HR image obtained from another camera. The results from each team will be evaluated in two ways as detailed below.

A dataset has been created using three different thermal cameras, in order to have real scenarios, with different conditions and objects. Each camera has a different resolution (low, mid, high). The technical information is shown below in Table 1 and illustrations from each camera depicted in Fig. 1.

Figure 1 - Examples from each camera.

A total of 1021 thermal images were simultaneously taken with each camera. All images are semi-paired. From all these images, 1001 are going to be published (951 for training and 50 for testing), the remaining 20 will be used in the challenge for validation.

PSNR and structural similarity (SSIM) measures are going to be computed over a small set of images left for evaluating the performance of the proposed solution. Two kinds of evaluations are going to be performed. For the first evaluation, a set of 10 downsampled and noisy images will be shared for traditional evaluation as shown in Fig. 2. For the second evaluation a set of 10 real images, from the MR dataset, will be shared. In this case, the obtained SR images will be compared with the corresponding real High Resolution semi-paired images, as shown in Fig. 3. For the second evaluation, HR images are going to be registered with the computed SR images in order to use PSNR & SSIM metrics.

Figure 2 - First evaluation diagram (x2 for low, x3 for mid, x4 for high).


Figure 3 - Second evaluation diagram*.

*Due to the semi-paired nature, for a fair comparison, the evaluation will be performed over a small region (50% of image size) centered in the image.

• Evaluation 1: will consist of three evaluations, one per each camera (x2 for low, x3 for mid and x4 for high). For each scale, noisy downsampled images from each camera will be provided to compute the corresponding SR (bicubic function will be used for downsampling and Gaussian noise** with mean=0 and sigma=10 will be added) from each camera will be provided to compute the corresponding SR, which will be evaluated as follows:
  where EVAL is PSNR & SSIM measures, xr corresponds to the SR scale and N is the number of validation images (10 per camera).
  
  ** On python for Gaussian noise, use np.random.normal(mean, sigma, img.shape)


• Evaluation 2: MR images as provided by the camera will be shared; this evaluation will be performed just for the x2 scale. The average evaluation value will be computed as follows:
  where EVAL is PSNR and SSIM measures, and N is the number of validation images (note that EVAL will be applied just on a part of the image to avoid regions without overlap due to the bias of the cameras).

The super-resolution (SR) results must be submitted in a zip file together with a short description of the proposed approach to <pbvs20.tisr.challenge@gmail.com>. The approach with the higher performance in most evaluation metrics will be the winner of the challenge.

• Registration open & datase released: December 10, 2019
• Evaluation images distributed: February 21, 2020
• Deadline for challenge & result submitted: February 28, 2020
• Winner announcement: June 14, 2020