The decline in bee population poses a significant threat to agricultural productivity and biodiversity, highlighting the need for advanced monitoring techniques. While deep learning techniques have shown remarkable potential for object detection tasks, their comparative performance have not been fully examined for the specific application of detecting and counting bees, especially in challenging environments where occlusion and small objects are common. This paper aims to conduct a performance analysis of deep learning-based object detection models for bee detection and counting to monitor hive conditions in Philippine bee farm. The study collected 3050 video clips, resulting in 58,000 frames, from which 5000 images with 11,775 annotations were pre-processed into the BukBee dataset. This dataset serves as the foundation for training and evaluating five advanced object detection models: SSD, Faster R-CNN, YOLOv4, EfficientDet, and CenterNet. The experimental results show that Yolov4 achieved the highest mAP score of 97.9 while SSD, Faster RCNN, EfficientDet, and CenterNet exhibited mAP scores of 82.3, 82.2, 73.3, and 79.7 respectively. The experiments reveal that the best performing model, Yolov4, demonstrated superior capabilities for bee detection and counting, even in challenging scenarios like occlusion and presence of small objects. The performance analysis presented in this study can be integrated into an IoT framework for real-time monitoring of bee hives, thereby promoting sustainable practices in bee farming.