TontonTremblay/doper2-hope-results

DOPER2 V2 Results on HOPE

Per-object DOPER2 V2 results on the BOP HOPE dataset. V2 uses the improved render config (HANDal-style framing, 2× distractors, 50% no-ground).

+ means above the 3PT-Pose-H3 SOTA aggregate; − means below. Differences are percentage points.

Summary

Method	Modality	AR	AR_MSSD	AR_MSPD	Time [s]
DOPER2 V2	RGB	60.4% (+0.49)	57.2% (-0.02)	63.6% (+0.99)	~0.01
3PT-Pose-H3 (SOTA)	RGB-D	59.9%	57.2%	62.6%	50.834

DOPER2 V2 average detection rate: 88.5% across all 28 objects.

Per-Object Results

Object	Name	AR_MSSD	vs 57.2%	AR_MSPD	vs 62.6%	AR_mean	vs 59.9%	Det. Rate	Diam. [mm]
000001	AlphabetSoup	52.9%	-4.31	64.4%	+1.84	58.7%	-1.23	80%	107.9
000002	BBQSauce	58.9%	+1.69	67.6%	+4.96	63.2%	+3.32	76%	153.9
000003	Butter	66.7%	+9.47	80.9%	+18.29	73.8%	+13.88	89%	115.0
000004	Cherries	50.5%	-6.70	69.0%	+6.40	59.8%	-0.15	90%	89.6
000005	ChocolatePudding	44.0%	-13.20	61.7%	-0.93	52.8%	-7.07	77%	98.0
000006	Cookies	69.6%	+12.40	55.6%	-7.00	62.6%	+2.70	100%	208.7
000007	Corn	45.2%	-11.95	57.0%	-5.60	51.1%	-8.78	82%	89.6
000008	CreamCheese	51.2%	-6.00	56.8%	-5.80	54.0%	-5.90	84%	115.6
000009	GreenBeans	64.0%	+6.80	58.0%	-4.60	61.0%	+1.10	100%	206.0
000010	GreenBeans2	41.7%	-15.49	54.3%	-8.31	48.0%	-11.90	80%	89.7
000011	Ketchup	61.0%	+3.80	67.0%	+4.40	64.0%	+4.10	80%	153.9
000012	MacaroniAndCheese	62.4%	+5.20	51.8%	-10.80	57.1%	-2.80	94%	207.1
000013	Mayo	70.3%	+13.13	72.3%	+9.73	71.3%	+11.43	100%	153.4
000014	Milk	89.6%	+32.40	80.0%	+17.40	84.8%	+24.90	100%	204.5
000015	Mushrooms	39.4%	-17.77	52.3%	-10.31	45.9%	-14.04	74%	75.7
000016	Mustard	72.0%	+14.80	73.2%	+10.60	72.6%	+12.70	94%	161.5
000017	OrangeJuice	42.4%	-14.80	38.0%	-24.60	40.2%	-19.70	72%	205.7
000018	Parmesan	60.3%	+3.13	72.0%	+9.40	66.2%	+6.27	83%	122.8
000019	Peaches	72.0%	+14.80	84.5%	+21.90	78.2%	+18.35	95%	89.2
000020	PeasAndCarrots	66.7%	+9.47	70.3%	+7.73	68.5%	+8.60	93%	89.9
000021	Pineapple	48.0%	-9.20	66.9%	+4.31	57.5%	-2.45	93%	89.2
000022	Popcorn	50.0%	-7.20	64.0%	+1.40	57.0%	-2.90	100%	152.4
000023	Raisins	51.2%	-5.95	53.2%	-9.35	52.2%	-7.65	82%	151.3
000024	SaladDressing	82.0%	+24.80	77.0%	+14.40	79.5%	+19.60	100%	151.3
000025	Spaghetti	64.9%	+7.66	58.6%	-4.03	61.7%	+1.81	100%	252.8
000026	TomatoSauce	28.9%	-28.34	47.4%	-15.17	38.1%	-21.76	69%	107.1
000027	Tuna	31.1%	-26.06	44.3%	-18.31	37.7%	-22.19	94%	76.1
000028	Yogurt	64.0%	+6.80	82.5%	+19.90	73.2%	+13.35	95%	82.9
	Average	57.2%	-0.02	63.6%	+0.99	60.4%	+0.49	88%
	3PT-Pose-H3 (SOTA)	57.2%		62.6%		59.9%		—

Inference Grids

Each grid shows detection + pose overlay on four data sources: Domain-Randomized synthetic, BOP PBR, Onboarding (real), and Validation (real). Green contour = prediction, Red contour = ground truth. Inset shows per-vertex error heatmap.

obj_000001 — AlphabetSoup

obj_000002 — BBQSauce

obj_000003 — Butter

obj_000004 — Cherries

obj_000005 — ChocolatePudding

obj_000006 — Cookies

obj_000007 — Corn

obj_000008 — CreamCheese

obj_000009 — GreenBeans

obj_000010 — GreenBeans2

obj_000011 — Ketchup

obj_000012 — MacaroniAndCheese

obj_000013 — Mayo

obj_000014 — Milk

obj_000015 — Mushrooms

obj_000016 — Mustard

obj_000017 — OrangeJuice

obj_000018 — Parmesan

obj_000019 — Peaches

obj_000020 — PeasAndCarrots

obj_000021 — Pineapple

obj_000022 — Popcorn

obj_000023 — Raisins

obj_000024 — SaladDressing

obj_000025 — Spaghetti

obj_000026 — TomatoSauce

obj_000027 — Tuna

obj_000028 — Yogurt

Training Pipeline

Each object follows a 5-stage pipeline:

1. Blend setup — PLY mesh → Blender file with 64 FPS keypoints
2. Render 10k — V2 config: fill_factor 0.15–0.6, up to 4 distractors, 50% no-ground
3. V1 Training — 120 epochs on synthetic data only
4. V4 Training — 15 epochs fine-tuning with BOP PBR data mixed in
5. V5 Training — 15 epochs fine-tuning with PBR + onboarding pseudo-labels

Architecture: DOPER2 (ConvNeXt-Tiny DINOv3 backbone, CenterNet det + mask + heatmap keypoints, 64 keypoints per object).

SOTA Reference

Rank	Date	Method	Type	Modality	AR	AR_MSSD	AR_MSPD	Time [s]
1	2025-10-01	3PT-Pose-H3 (f.k.a IPT)	Single	RGB-D	59.9%	57.2%	62.6%	50.834