UnitBox is a novel object detection network that improves bounding box prediction by using an Intersection over Union (IoU) loss function. Traditional methods assume bounding box coordinates are independent, but this leads to less accurate localization. UnitBox regresses the four bounds of a predicted box as a whole unit, using the IoU loss to enforce maximal overlap between the predicted and ground truth boxes. This approach results in more accurate and efficient localization, and is robust to objects of varied shapes and scales. UnitBox is applied to face detection and achieves the best performance on the FDDB benchmark. The IoU loss layer is introduced to address the limitations of the $ \ell_{2} $ loss. Unlike $ \ell_{2} $, which treats coordinates as independent variables, IoU loss considers the bounding box as a single unit, leading to more accurate predictions. The IoU loss is also more robust to scale variations, allowing UnitBox to handle objects of different sizes effectively. UnitBox is based on a fully convolutional network architecture, which allows it to predict object bounds and classification scores directly from feature maps. It uses a novel IoU loss layer for bounding box prediction, which jointly regresses all four bounds as a unit. This approach leads to faster convergence and more accurate localization. UnitBox is trained with multi-scale objects and tested on single-scale images, making it efficient and effective. The UnitBox network is trained on the WiderFace dataset and fine-tuned using mini-batch SGD. It outperforms the $ \ell_{2} $ loss in terms of convergence speed and detection accuracy. The UnitBox achieves the best performance on the FDDB benchmark, and is efficient enough to run at about 12 fps on VGA-sized images, making it suitable for real-time detection systems. The IoU loss and UnitBox are shown to be effective for object detection and localization tasks.UnitBox is a novel object detection network that improves bounding box prediction by using an Intersection over Union (IoU) loss function. Traditional methods assume bounding box coordinates are independent, but this leads to less accurate localization. UnitBox regresses the four bounds of a predicted box as a whole unit, using the IoU loss to enforce maximal overlap between the predicted and ground truth boxes. This approach results in more accurate and efficient localization, and is robust to objects of varied shapes and scales. UnitBox is applied to face detection and achieves the best performance on the FDDB benchmark. The IoU loss layer is introduced to address the limitations of the $ \ell_{2} $ loss. Unlike $ \ell_{2} $, which treats coordinates as independent variables, IoU loss considers the bounding box as a single unit, leading to more accurate predictions. The IoU loss is also more robust to scale variations, allowing UnitBox to handle objects of different sizes effectively. UnitBox is based on a fully convolutional network architecture, which allows it to predict object bounds and classification scores directly from feature maps. It uses a novel IoU loss layer for bounding box prediction, which jointly regresses all four bounds as a unit. This approach leads to faster convergence and more accurate localization. UnitBox is trained with multi-scale objects and tested on single-scale images, making it efficient and effective. The UnitBox network is trained on the WiderFace dataset and fine-tuned using mini-batch SGD. It outperforms the $ \ell_{2} $ loss in terms of convergence speed and detection accuracy. The UnitBox achieves the best performance on the FDDB benchmark, and is efficient enough to run at about 12 fps on VGA-sized images, making it suitable for real-time detection systems. The IoU loss and UnitBox are shown to be effective for object detection and localization tasks.