Two thresholds at 0.25 decide which of Grounding-Dino's 900 default boxes survive

The scoring loop is the part to read before budgeting a labeling run. The guide says the model produces 900 boxes by default, scores them against the input words, keeps a box when its highest similarity exceeds `box_threshold`, and takes the words whose similarity clears `text_threshold` as that box's label ^[10]. Both thresholds accept 0 to 1 and both default to 0.25 ^[15]. Tuning comes down to the query string and those two numbers ^[18], because the schema carries no resizing, non-maximum-suppression or batch controls ^[8].

Scoring happens against tokens. One word can split into several tokens, and the guide says the number of words in a sentence is not necessarily the number of text tokens ^[11]. A description combining several concepts can therefore behave differently from the same categories sent separately ^[14]. The label-extraction step is also how you end up with labels that do not match the phrase you typed, alongside missed detections and duplicate or weak boxes ^[12].

The hosted query field is described as "Comma seperated names of the objects" ^[9], typo included. The guide calls the gap between that and the upstream README's periods an interface mismatch worth testing ^[9]. I would send one image twice, commas once and periods once, and compare which boxes clear 0.25 ^[15].

Both accuracy figures come from the underlying project and not the hosted build: 52.5 AP zero-shot on COCO with no COCO training data ^[4], and 63.0 AP after COCO fine-tuning ^[5]. The 10.5-point difference is what fitting one fixed class list buys ^[6]. For the zero-shot number to say anything about your run, your prompts would have to behave like COCO's category names on images like COCO's, and a query naming a red safety helmet, a forklift and a damaged package ^[16] is measuring something else.

The reason to adopt it anyway is that the class list is a string. You can add a category no fixed detector carries without retraining anything, and the guide says to treat what comes back as proposals that require review, especially for production datasets ^[16]. For inspection prototypes it says threshold tuning and human validation stay necessary where the consequences are high ^[19].

Downstream, boxes are what you get. The project's workflows feed them to Grounded SAM and Grounded SAM 2 for masks, and to Stable Diffusion and GLIGEN as spatial grounding for edits; the detector does not return pixel masks or tracking trajectories itself ^[17]. Depth, pose, identity, attributes with guaranteed reliability and consistency across video frames are all beyond it ^[13]. It will put a box on a forklift without telling you which forklift ^[13]. Small, occluded, unusual, abstract or visually ambiguous objects can fail ^[14].