PyCharm blog walks through fine-tuning YOLO12, YOLO26, and RF-DETR detectors on RF100-VL off-distribution datasets

The PyCharm data science blog published a practical tutorial by Arina Belova that follows up an earlier theoretical post on the architectures of three SOTA object detection families. The new post focuses on actually using the models: loading pretrained checkpoints, reproducing COCO baselines, and beginning to fine-tune on RF100-VL. Six checkpoints were used in total - two sizes each of YOLO12, YOLO26, and RF-DETR - sourced from the original authors' repo, the Ultralytics PyPI package, and the Roboflow PyPI package.

Before touching off-distribution data, the author evaluated all six checkpoints on the full 5,000-image COCO val2017 split to confirm the numbers reported in the prior post. RF-DETR Base led on mAP50-95 at 0.5325 with 32.1M parameters, narrowly ahead of YOLOv12-M at 0.5259 with 19.6M parameters. YOLO26-N was the fastest model overall at 12.3 ms latency, benefiting from its NMS-free design, while YOLOv12-N was markedly slower at 23.9 ms despite a similar accuracy profile. RF-DETR Nano, despite the name, has roughly 30M parameters - more than YOLO26-M - but stayed well-optimized at 12.4 ms latency.

The latency numbers diverge from the figures in the models' white papers, and the author is upfront about two reasons. First, the hardware differed from the NVIDIA T4 GPU that has become the de facto benchmark standard in object detection. Second, the models were run in their native frameworks rather than compiled to TensorRT, which would otherwise fuse layers, pick the fastest kernels for the target GPU, and optionally run in reduced precision. The TensorRT engine is tied to one GPU and adds a build step, so it does not represent out-of-the-box performance. Accuracy told a different story, with mAP50-95 falling within reasonable noise bounds of the reported figures.

Pretrained detectors are trained on COCO, a corpus of roughly 118,000 images covering 80 everyday object categories. Real-world deployment targets such as damaged industrial cables, bone fractures on X-rays, or densely stacked retail shelf items are typically both out-of-vocabulary and out-of-distribution from COCO. "Bone fracture" is not one of the 80 classes, so the model has no output category for it at all, and the off-distribution visual style of X-rays, industrial close-ups, or heavily occluded shelf scenes differs drastically from consumer photos in texture, viewpoint, and object density.

For the off-distribution evaluation the author selected RF100-VL, a large-scale collection of 100 multimodal datasets whose concepts were deliberately chosen to be rare in object detection pretraining data. The provided extract ends at the point of introducing these datasets as the targets for fine-tuning, so the actual fine-tuning procedure and post-tuning accuracy numbers are not available here; engineers reading the tutorial should expect those results further down the post.