Player not loading? Watch on YouTube
This tutorial compares SAM3 with Grounding DINO + SAM2 for kidney histology segmentation. Grounding DINO detects bounding boxes from text; SAM2 turns those boxes into masks. SAM3 accepts text and produces masks directly. The speaker tests both approaches in the same annotation tool, including a fine-tuned DINO checkpoint.
With the prompt "glomerulus" and a threshold of 0.15, SAM3 finds no objects in the example. The base DINO + SAM2 pipeline detects two glomeruli, while the fine-tuned pipeline detects several with cleaner masks. The speaker reports training on 24 annotated images in under five minutes on his home GPU, with a validation F1 of 0.70. These results describe his experiment, not a general benchmark.
Changing the prompt to "small circular objects" improves SAM3's results. A later mitochondria test follows the same pattern: "mitochondria" returns nothing, while "elongated oval organelle" produces detections, including false positives.
The setup section explains how to request SAM3.1 access through Hugging Face and download weights to run models locally. The speaker quotes approximately 16 GB of VRAM for SAM3. He also demonstrates SAM2's click-based mask correction; the SAM3 backend disables that control, and he acknowledges that his broader claim about missing click support is an assumption.
