Point annotations offer a lightweight alternative to bounding boxes for object detection, particularly in aerial wildlife monitoring where objects are small, dense, and precise extents are unnecessary. However, most detectors are designed for box supervision. We introduce DINOKey, a transformer-based detector that operates directly on 2D keypoints by adapting the DINO framework to predict point coordinates. The model combines L1 regression loss, focal loss, and average Hausdorff distance to enforce both local accuracy and global consistency. On an aerial elephant dataset, DINOKey achieves a mAP of 0.786. Ablations show the Hausdorff term is critical in dense scenes, while focal loss improves stability under class imbalance. Results demonstrate that point-based transformer detectors are an effective alternative to box-based methods for small-object, high-density settings.Point annotations offer a lightweight alternative to bounding boxes for object detection, particularly in aerial wildlife monitoring where objects are small, dense, and precise extents are unnecessary. However, most detectors are designed for box supervision. We introduce DINOKey, a transformer-based detector that operates directly on 2D keypoints by adapting the DINO framework to predict point coordinates. The model combines L1 regression loss, focal loss, and average Hausdorff distance to enforce both local accuracy and global consistency. On an aerial elephant dataset, DINOKey achieves a mAP of 0.786. Ablations show the Hausdorff term is critical in dense scenes, while focal loss improves stability under class imbalance. Results demonstrate that point-based transformer detectors are an effective alternative to box-based methods for small-object, high-density settings.