gavinlaw/multimodal-understanding-lvis-scale-outputs

Object presence, scale and negative-evidence diagnostic

Frozen Qwen3-VL-4B, InternVL3.5-4B and GroundingDINO-tiny outputs on a deliberately size-stratified subset of LVIS v1 validation. This is a diagnostic, not an official LVIS benchmark score, trained model, novel search method or proof of generalization. See summary.json for the actual result, including failed gates. No VLM parameters were trained.

16 categories and 640 distinct image IDs/RGB hashes were selected before inference. Local train/dev/holdout sizes are 320/160/160. Only train/dev were evaluated. Complete-positive categories exclude not_exhaustive_category_ids; negatives require explicit neg_category_ids. Unannotated categories are not treated as negative. Each category includes small positive, explicit negative and large positive controls. Selection intentionally favors small objects and is not representative of all LVIS images. COCO/LVIS overlap with pretrained models is possible.

Whole low/high images have long edges 448/1792. High InternVL uses native dynamic patches with max16; Qwen uses max_pixels=1792 squared. Sixteen overlapping 1/3-width by 1/3-height windows use long edge 448. Each view is scored independently with the same fixed category definition. Scan uses the maximum yes score. High-long includes up to 768 generated reasoning tokens and a separate uniform yes/no readout; both calls count toward costs. Oracle uses a GT-box context crop only for diagnostic positive-image recognition and is not a deployable action. GroundingDINO uses its official default processor and category name query.

Yes/no conditional scores are not calibrated probabilities. Absolute answer-token mass is retained. Each ordinary condition must have binary mass at least 0.5 in at least 95% of calls. Thresholds for each condition are selected only on train negatives at empirical FPR at most 5%; all dev comparisons use frozen thresholds. The candidate prerequisite requires both VLMs to improve small-object recall over high and high-long by at least 10 percentage points with paired 95% CI lower bound above0, dev FPR at most 10%, and oracle small recall at least 80%. Passing this gate would still not establish a new method or efficient policy. Holdout remains unrun. Independent image bootstrap with category stratification uses 2,000 replicates. These exploratory screening thresholds are not a sequential coverage guarantee.

Costs include prefill tokens, actual visual tokens, all generation/readout calls, preprocessing and batch throughput time. Batch time divided by batch size is not single-request latency. Detector elapsed time is reported separately. All 20,256 calls are validated against input hashes and complete manifests.

Original LVIS metadata: https://www.lvisdataset.org/ . Source images are hosted by COCO and retain their individual licenses, referenced in sources.jsonl.gz. This release includes no original images, crops, segmentation masks or bbox arrays. Our code and generated measurements are CC BY 4.0 to the extent we hold rights; this does not relicense source images, annotations or pretrained models. Download source data under its original terms.

Reproduce on persistent compute storage: run source/audit_lvis_resources.py --root /external/lvis, then prepare_lvis_scale.py with the same root. This downloads the pinned annotation archive and source images, verifies 640 distinct RGB hashes, and creates inputs and separate labels. Run run_lvis_scale.py --root /external/lvis/scale-evaluation --model vl|intern --shard <0..3> --deadline-unix on explicitly allocated GPUs. Run run_lvis_detector.py after allocating a free GPU. Run validate_lvis_scale.py, then analyze_lvis_scale.py --root --output /summary.json. Heavy inputs stay on the compute host. Relative reconstruction paths are published; hashes of original inference requests include original absolute paths as documented in manifest.json.

Related methods already cover active search, local zoom, calibration and absence: ROAM/VLN-NF (https://arxiv.org/abs/2604.10533), Explore Until Confident (https://arxiv.org/abs/2403.15941), CVSearch (https://arxiv.org/abs/2605.23655), and VPS (https://arxiv.org/abs/2609.37002). This release does not claim those components as new.