Annotated HDR dataset

Supervised object detection requires large scale  annotated datasets. However, creating large HDR annotated datasets is a major challenge due to the following reasons:

  • There are no large scale HDR datasets publicly available.
  • Existing annotation tools do not support HDR images (so there is no way to open these files)
  • Video annotation is highly time consuming and resource intensive.

Therefore, to mitigate the above-mentioned challenges, we first shortlisted a set of seven HDR video sequences. From the video sequences, we carefully curated a total of  4080 images which contain some images from Google HDR dataset. The HDR images were then tone-mapped using a scene reproduction operator to create 8-bit representations of the HDR images. The tone-mapped images were then used to annotate the HDR dataset using the “labelimg” annotation tool. The tone-mapped images were discarded to produce a clean dataset with HDR images along with the annotations. Finally, the annotations were converted to COCO JSON format for ease of training and testing.