Negation as Excluded Volume in Open-Vocabulary Detectors

Vision-language detectors find the boxes that match a text description. They fail on negation in a specific, repeatable way. Asked for "a dog not led by a rope," they return the dog on the leash, because mentioning "rope" adds evidence for every region that contains one. Prior work traces this failure to two causes. Negation is rare in training text: under 0.1% of web captions contain a negation word. And the matching score behaves like a sum of concept evidence, so a negated concept can only add to the score, never subtract. This note proposes treating the score a detector assigns to a box like a physical density with a capacity limit. A description and its negated counterpart then compete for the same finite room in a box: whatever the box gives to "dog led by a rope" is no longer available to "dog not led by a rope."

The constraint comes from volume-filling models in transport physics, where particles cannot pile up beyond a site's capacity and the flux into a site shrinks with its remaining headroom. Its practical value here is about data. If the score for the negated description is defined as the headroom the affirmative description leaves behind, then every ordinary affirmative grounding example also supervises negation. The scarcity of negation data then matters much less. The central bet is that a capacity-coupled detector trained with little or no negation text can match detectors fine-tuned on large synthetic negation corpora.

Core Claim

Consider a box $b$ and a description of the form "head concept $A$ with modifier $M$ negated," for example $A$ = dog and $M$ = led by a rope. We replace the detector's single free score with a two-species occupancy that has a dual bound:

$$ s(b, A\wedge M) = \sigma(h_A(b))\,\sigma(h_{M|A}(b)), \qquad s(b, A\wedge\neg M) = \sigma(h_A(b))\,\bigl(1-\sigma(h_{M|A}(b))\bigr). $$

Here $\sigma$ is the logistic function. $h_A(b)$ is the detector's logit that box $b$ contains $A$. $h_{M|A}(b)$ is a logit for whether that instance of $A$ carries modifier $M$. Both are read from the same fused cross-attention features. Two consequences follow by construction. First, the two scores sum exactly to $\sigma(h_A(b)) \le 1$. Contradictory descriptions therefore cannot both claim the same box, which is the discrete single-occupancy rule and the $u+m\le 1$ constraint of the volume-filling models. Second, $h_{M|A}$ is trained by every affirmative example that mentions $M$. A negated query is answered with parameters that never needed negated text.

This is the exact, closed-form dual bound that FluxNet leaves open for neural time-stepping. FluxNet bounds transport between grid cells by the receiver's headroom $(1-\rho/\rho_{\max})$, but its dual-bounded head relies on auxiliary consistency losses. The guarantee is easy here because the transported quantity is a per-box scalar rather than a field on a stencil. We do not claim progress on FluxNet's PDE problem. The idea we borrow is narrower: let headroom, rather than an unbounded rate, determine what can still flow in.

Where This Sits

Three bodies of work motivate the proposal. The scarcity numbers come from What "Not" to Detect and Grounding the "Not": 0.08% of LAION-400M captions contain negation. NegVQA reports that only 1.1% of LLaVA dialogue turns contain "not," and that scaling does not close the gap: Qwen2-VL-72B falls from 92.2% to 72.7% on negated questions. The mechanism comes from Similarity Is Not Logic, which shows that dual-encoder dot products act as bag-of-concepts aggregators in which a mentioned absent concept adds positive mass. The detection failure comes from Described Object Detection. That paper builds the $D^3$ benchmark, in which about a quarter of descriptions express absence. It shows that referring-expression models err 100% of the time on absent targets, and that a post-hoc verifier cuts this only to 35.6%.

Two lines of work supply method and also set constraints. NegToMe with deep LoRA (What "Not" to Detect) merges negation tokens with a parser, boosts their embeddings by a fixed factor $\beta$, and fine-tunes the last three cross-attention layers of Grounding DINO on the 91,110-caption CoVAND set. It is our strongest competitor and our training substrate. We reuse its parser, LoRA placement, and backbone. We differ in two ways: we replace the unconstrained boost with a bounded, coupled score, and we ask how much negation data is actually needed. LCSE (Similarity Is Not Logic) already computes negation as $1-p$ on calibrated, LLM-factored retrieval scores. Our product form is its natural box-level analogue, so we treat a box-level LCSE as a baseline rather than claiming the algebra as new.

What is new relative to LCSE is the setting and the training. LCSE edits global image scores after the fact, using a sigmoid fitted on held-out data. We place the bound inside a dense detector and train $h_{M|A}$ end to end, conditioned on the instance of $A$. A rope detected elsewhere in the image therefore should not veto the dog.

The data-versus-architecture dispute constrains the design. Know "No" Better argues that fine-tuning on mined negation data fixes CLIP without hurting ImageNet. Seeing What's Not There argues that hard-negative fine-tuning causes forgetting and that training-free projection works better. Our affirmative-only training regime partly sidesteps the dispute: it never shows the model synthetic negatives. It also turns the dispute into a measurable question, namely whether negation data adds anything once the coupling is in place.

Research Program

We build the capacity head on Grounding DINO-B. The parser that NegToMe uses splits each description into a head span and an optional modifier span. If the parser finds no negation, the box score is simply $\sigma(h_A)\sigma(h_{M|A})$, or $\sigma(h_A)$ when there is no modifier. If it finds negation, the score is the headroom form above. Both logits come from small heads on the final fused query features, and LoRA is placed exactly as in NegToMe.

Training uses grounding data in three regimes. In the affirmative-only regime, the model sees the positive half of CoVAND plus standard referring and grounding data, with no negated captions. The partial regime adds 10% of CoVAND's negated captions. The full regime uses all of CoVAND. The question is how quickly negation accuracy saturates as negation data is added, compared with an unconstrained head.

We then analyze where the coupling helps and where it cannot. Three cases are most likely to break it. In the first, the modifier is not attached to the head instance, as in "a room with no people." In the second, the negation is implicit, as in "an empty room." In the third, the negated span is itself the head, as in "not a dog," which has no positive head to take headroom from. We will report these as separate subsets of $D^3$ and OVDEval rather than averaging them away.

Baselines

The naive reference point is off-the-shelf Grounding DINO-B, scored directly on the full negated description. Described Object Detection reports 20.7 mAP for Grounding DINO on intra-scenario $D^3$. We will rerun it under our evaluation code so that every method is scored the same way.

The strongest existing method is NegToMe with deep LoRA, trained on full CoVAND. Its paper reports gains of +10.8 class-ignored NMS-AP and a 19.1% reduction in false-positive rate on the OVDEval negation subset, plus +6.6 mAP on $D^3$. The file gives only these deltas, not absolute values, so we will reproduce the method. We give it the same backbone, the same three LoRA layers, the same parser, and the same number of hyperparameter trials, including a sweep over $\beta$. Beating it means matching its OVDEval NMS-AP in the 10%-data regime, and exceeding it in the full-data regime on both NMS-AP and false-positive rate on absence descriptions. Doing either while also degrading affirmative descriptions on $D^3$ by less than it does would count as well.

The verifier baseline is OFA-DOD from Described Object Detection, which reports 21.6 mAP intra-scenario and a 35.6% false-positive error on absent targets. We compare against those reported figures. Separately, we train a verifier of matched size on top of Grounding DINO-B so that the backbones are equal.

The post-hoc logic baseline is box-level LCSE in the style of Similarity Is Not Logic. Frozen Grounding DINO separately scores "dog" and "dog led by a rope." The two scores are calibrated with a sigmoid fitted on held-out data and then combined as $p_A(1-p_{A\wedge M})$, using the same parser. The file has no reported result for LCSE on detection benchmarks; its 85.5% on FACTOR-Bench is a retrieval result.

The key ablation is an unconstrained twin. It has the same two heads and the same data, but the negated description receives its own free logit instead of the headroom term. This isolates the coupling from the extra parameters and the parser.

Why not just…?

Why not just apply LCSE's $1-p$ to box scores and skip training? This is the most serious threat, because the algebra is the same. Our bet is that a frozen detector's score for "dog led by a rope" is badly calibrated as a conditional on the dog instance. The rope gets localized on its own, and a sigmoid fitted on held-out data cannot repair errors that depend on where things are in the image. The box-level LCSE baseline tests this directly. If it comes within one NMS-AP point of the trained coupled head on OVDEval and $D^3$ absence descriptions, the honest conclusion is that the constraint, not the training, does the work. The note would then become a recommendation for post-hoc box logic.

Why not just generate more synthetic negation data, as CoVAND and NegationCLIP do? That route works: NegToMe's gains and NegationCLIP's 62.95% versus 57.27% on NegRefCOCOg (Know "No" Better) show it. But each pipeline needs GPT-4o generation plus verification by a multimodal model, and the forgetting concern raised in Seeing What's Not There remains unresolved. The data-budget sweep settles the question. If the unconstrained twin and NegToMe close the gap as soon as 10% of CoVAND is added, the coupling offers only data efficiency. That is still useful in low-resource languages, where Disparities in Negation Understanding Across Languages shows the same bias, but it is a smaller claim.

Why not just run class-ignored NMS or a verifier at inference? Class-ignored NMS removes double detections only after they occur, and it does not decide which of two contradictory labels is correct. The OFA-DOD verifier still leaves 35.6% false-positive error on absent targets. Our construction makes the two labels' scores sum to the head score, so the decision is forced inside the model. We will report the double-detection rate before NMS to show whether this holds in practice.

Why not just scale the model? NegVQA documents a U-shaped relationship between scale and negation accuracy, and a 19.5-point drop even at 72B parameters. We do not test scaling ourselves, because detectors of that size are not available in the material.

Proposed Experiments

All runs use Grounding DINO-B trained on CoVAND in the three data regimes. We evaluate on the OVDEval negation subset, reporting class-ignored NMS-AP and false-positive rate. On $D^3$, we report mAP split into presence and absence descriptions, plus the pre-NMS double-detection rate. For zero-shot transfer, we use FG-CXR, where NegToMe reports a rise from 54.8% to 62.5%.

The decisive check is a single comparison: the capacity head trained affirmative-only, against NegToMe trained on full CoVAND, on absence descriptions in OVDEval and $D^3$. If affirmative-only training reaches at least NegToMe's NMS-AP, the claim that coupling converts affirmative supervision into negation competence holds. If it falls short, we move to the second check: whether the capacity head beats the unconstrained twin at every data budget. A gap that shrinks as data grows would mean the constraint is a data-efficiency device and nothing more.

The third check is the failure-case breakdown: unattached modifiers, implicit negation, and negated heads. It shows where a headroom formulation stops applying and where learned scope detection would have to replace the parser.