LocateAnything

Full benchmark tables, ablation studies, and qualitative visualizations. For a high-level overview, see the project README.

---

Table of Contents

• Headline Numbers
• 1. Common Object Detection — LVIS & COCO
• 2. Dense Object Detection — Dense200 & VisDrone
• 3. Document Layout & OCR
• 4. GUI Grounding — ScreenSpot-Pro
• 5. Referring Expression Comprehension
• 6. Pointing
• 7. Ablation Study
• 8. Qualitative Results

---

Headline Numbers

	LocateAnything-3B
Throughput on a single H100	12.7 BPS
Speedup vs. Qwen3-VL (1.1 BPS)	≈10×
Speedup vs. Rex-Omni (5.0 BPS)	≈2.5×
LVIS F1@Mean	50.7 (+3.8 over Rex-Omni)
COCO F1@Mean	54.7 (+1.8 over Rex-Omni)
Dense200 F1@Mean	58.7
DocLayNet F1@Mean	76.8
M6Doc F1@Mean	70.1 (+14.5 over Rex-Omni)
TotalText F1@Mean	43.3
ScreenSpot-Pro Avg	60.3 (SOTA)
HumanRef F1@0.95	68.8 (SOTA)
RefCOCOg val F1@Mean	76.7 (SOTA)
Pointing — best on all 7 benchmarks	✅

---

1. Common Object Detection — LVIS & COCO

Method	BPS	LVIS F1@0.5	LVIS F1@0.95	LVIS F1@Mean	COCO F1@0.5	COCO F1@0.95	COCO F1@Mean
Grounding DINO-Swin-T	—	47.7	22.7	38.8	69.8	23.0	56.6
DeepSeek-VL2-Small	—	56.2	21.0	41.8	60.9	14.9	45.9
Qwen3-VL-4B	1.1	59.8	20.0	43.5	63.0	14.2	46.1
Qwen3-VL-8B	1.0	61.5	20.2	44.8	62.8	14.0	45.7
SEED1.5-VL	—	65.6	19.5	46.7	71.3	14.3	51.4
Rex-Omni-3B	5.0	64.3	20.7	46.9	72.0	15.9	52.9
LocateAnything-3B	12.7	62.3	31.1	50.7	70.1	19.3	54.7

LocateAnything improves the mean F1 by +3.8% on LVIS and +1.8% on COCO over Rex-Omni at identical model size, with particularly strong gains at high IoU thresholds (31.1 vs. 20.7 at IoU=0.95 on LVIS).

---

2. Dense Object Detection — Dense200 & VisDrone

Method	Dense200 F1@0.5	Dense200 F1@0.95	Dense200 F1@Mean	VisDrone F1@0.5	VisDrone F1@0.95	VisDrone F1@Mean
Grounding DINO-Swin-T	36.9	19.7	33.1	55.2	3.9	38.5
Qwen3-VL-4B	17.5	2.4	12.5	42.3	1.4	26.0
SEED1.5-VL	76.9	5.3	53.2	55.9	0.6	27.4
Rex-Omni-3B	78.4	10.3	58.3	61.6	1.5	35.8
LocateAnything-3B	74.0	18.5	58.7	63.0	3.2	39.9

LocateAnything achieves 58.7 and 39.9 mean F1 on Dense200 and VisDrone respectively, substantially outperforming Rex-Omni (58.3 / 35.8), demonstrating superior boundary delineation in heavily overlapping scenes.

---

3. Document Layout & OCR

Method	DocLayNet F1@0.5	DocLayNet F1@0.95	DocLayNet F1@Mean	M6Doc F1@0.5	M6Doc F1@0.95	M6Doc F1@Mean	TotalText F1@0.5	TotalText F1@0.95	TotalText F1@Mean
DocLayout-YOLO	91.2	52.1	81.1	—	—	—	—	—	—
Qwen3-VL-4B	60.8	8.2	37.2	30.6	4.9	19.0	55.4	3.6	36.1
Qwen3-VL-8B	54.7	6.7	34.1	37.2	4.9	22.7	59.4	2.7	37.3
Rex-Omni-3B	89.5	28.4	70.7	76.3	18.7	55.6	56.6	3.9	40.6
LocateAnything-3B	91.1	35.8	76.8	90.6	25.8	70.1	58.9	5.1	43.3

New standards on document understanding: 76.8 and 70.1 mean F1 on DocLayNet and M6Doc respectively, outperforming Rex-Omni by substantial margins (+6.1 / +14.5). On TotalText OCR, 43.3 mean F1 surpasses all compared methods.

---

4. GUI Grounding — ScreenSpot-Pro

Method	Dev T/I	Creative T/I	CAD T/I	Science T/I	Office T/I	OS T/I	Avg
Rex-Omni-3B	61.7 / 9.7	52.5 / 12.6	22.3 / 9.4	59.0 / 26.4	63.3 / 28.3	24.1 / 15.7	36.8
GUI-Owl-7B	76.6 / 31.0	59.6 / 27.3	64.5 / 21.9	79.1 / 37.3	77.4 / 39.6	59.8 / 33.7	54.9
MAI-UI-2B	76.6 / 32.4	69.2 / 21.7	61.4 / 23.4	81.2 / 34.5	85.9 / 39.6	68.2 / 41.6	57.4
GUI-Owl-32B	84.4 / 39.3	65.2 / 18.2	62.4 / 28.1	82.6 / 39.1	81.4 / 39.6	70.1 / 36.0	58.0
LocateAnything-3B	70.8 / 50.3	60.1 / 46.9	57.9 / 40.6	69.4 / 58.2	77.2 / 69.8	65.4 / 43.8	60.3

SOTA average 60.3, surpassing generalist VLMs like Qwen3-VL-30B-A3B and specialized models such as GUI-Owl-32B, with particularly strong performance on icon-based queries.

---

5. Referring Expression Comprehension

Method	HumanRef F1@0.5	HumanRef F1@0.95	HumanRef F1@Mean	RefCOCOg val F1@Mean	RefCOCOg test F1@Mean
Qwen3-VL-4B	77.7	54.9	71.1	74.7	74.6
Qwen3-VL-8B	78.6	55.7	72.0	74.9	75.2
SEED1.5-VL	88.2	60.0	81.6	71.9	73.2
Rex-Omni-3B	85.4	65.4	79.9	73.6	74.3
LocateAnything-3B	82.9	68.8	78.7	76.7	77.6

Seamlessly aligns nuanced human intents with visual regions, achieving 78.7 mean F1 on HumanRef and remaining highly competitive on RefCOCOg against top-tier models.

---

6. Pointing

Method	COCO	LVIS	Dense200	VisDrone	HumanRef	RefCOCOg val	RefCOCOg test
Molmo-7B-D	77.3	40.3	33.1	29.2	70.0	83.7	83.6
SEED1.5-VL	78.2	70.7	72.1	56.7	83.1	83.6	84.2
Rex-Omni-3B	80.5	70.8	82.5	58.9	83.8	84.7	85.1
LocateAnything-3B	83.9	76.6	87.6	60.4	84.7	91.3	91.0

Best across all 7 point-based grounding benchmarks.

---

7. Ablation Study

Design choices and decoding efficiency on COCO.

• (a) Coordinate representation. PBD (Slow Mode) achieves the highest F1 of 52.1, confirming that box-aligned atomic formulation provides stronger supervision than 1D serialization.
• (b) MTP formulation. PBD dramatically outpaces structure-agnostic MTP methods (16.9 BPS vs. 5.5 BPS for SDLM-B6) while improving F1.
• (c) Decoding modes. Joint training pushes Slow Mode to 52.1 F1; Hybrid Mode preserves most speed gains (13.2 BPS at 51.6 F1).

Decoding Mode Comparison

Joint dual-formulation training successfully pushes the Slow Mode upper bound from 50.1 to 52.1 F1. Hybrid Mode seamlessly resolves the speed–accuracy trade-off, achieving robust high-precision localization while preserving most speed gains.

Box Ordering & Decoding Throughput

• Left: X-Y corner-order sorting yields the highest F1 among four spatial ordering strategies.
• Right: as the number of target boxes grows from 20 to 300, NTP methods suffer severe latency bottlenecks, while Parallel Box Decoding achieves a 2×–6× speedup, scaling throughput from 12 BPS to ~25 BPS in dense scenes.

---

8. Qualitative Results

High-quality grounding in the wild across documents, GUIs, and natural images.

_{Qualitative visualizations of dense and high-precision box predictions across diverse resolutions and categories.}

---

Citation

@article{wang2025locateanything,
  title   = {LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding},
  author  = {Shihao Wang and Shilong Liu and Yuanguo Kuang and Xinyu Wei and
             Yangzhou Liu and Zhiqi Li and Yunze Man and Guo Chen and
             Andrew Tao and Guilin Liu and Jan Kautz and Lei Zhang and Zhiding Yu},
  journal = {arXiv preprint arXiv:25XX.XXXXX},
  year    = {2026},
}