pallet-dope-cropaug-truncation (DOPE, crop-aug, truncation-best)

νŒ”λ ˆνŠΈ 6D 포즈 μΆ”μ •μš© DOPE λͺ¨λΈ. crop + padding μ¦κ°•μœΌλ‘œ fine-tune ν•˜μ—¬, μ‹€μ œ 카메라가 쒌우둜 νŒ¨λ‹ν•˜λ©° νŒ”λ ˆνŠΈκ°€ ν™”λ©΄ λ°–μœΌλ‘œ μž˜λ¦¬λŠ”(truncation) μƒν™©μ—μ„œ κ°•κ±΄ν•˜κ²Œ λ™μž‘ν•˜λŠ” 것이 νŠΉμ§•.

  • best checkpoint: final_net_epoch_0180.pth (λˆ„μ  epoch 180)
  • convention: camera-facing 0123 (0~3 μ•žλ©΄, {0,1,4,5}=μœ„ / {2,3,6,7}=μ•„λž˜, 8=centroid)
  • backbone: VGG-19, 9 belief maps + 16 affinity fields
  • input: 448Γ—448, sigma=4.0

μ™œ 이 λͺ¨λΈμΈκ°€ (truncation best)

baseline DOPE λŒ€λΉ„ real truncation ν™˜κ²½μ—μ„œ:

metric              baseline    crop-aug (this)
─────────────────────────────────────────────
detection rate      13%         94%
PnP success         23%         99%

YOLO-pose 의 crop+padding 증강 방식을 DOPE 둜 μ΄μ‹ν•˜μ—¬ 얻은 κ²°κ³Ό.

ν•™μŠ΅ μ„€μ •

net_path  : dope_cropaug_ft_s1/final_net_epoch_0150.pth (μ΄μ–΄μ„œ fine-tune)
epochs    : 180 (λˆ„μ  λͺ©ν‘œ, base 150 + μΆ”κ°€ 30)
lr        : 5e-5
batchsize : 4
imagesize : 448
sigma     : 4.0
seed      : 3709
data      : capturepallet{02~09,cad} + capturenight{04~09}
            + forklift_raw + truncation_crops_dope/ft_real (manual GT)

전체 ν•™μŠ΅ μΈμžλŠ” header.txt μ°Έμ‘°.

평가 (synthetic val, 200 frames)

eval_summary.json:

PCK@3px   : 0.224     PnP success : 0.555
PCK@5px   : 0.277     reproj median: 164.6 px*
PCK@10px  : 0.416     volume ratio median: 0.836

* reproj μˆ˜μΉ˜λŠ” evaluate_on_val 의 convention 뢈일치 λ²„κ·Έλ‘œ κ³ΌλŒ€ν‰κ°€λ¨ (μ‹€μ œ reproj λŠ” ν•œ 자리수 px). κ²€μΆœλ₯ Β·PnP success λŠ” μ‹ λ’° κ°€λŠ₯. truncation 강건성은 real ν™˜κ²½(μœ„ ν‘œ)μ—μ„œ 확인.

μ‚¬μš©

import torch
# DOPE DreamNetwork μ •μ˜λŠ” Deep_Object_Pose/common/models.py μ°Έμ‘°
state = torch.load("final_net_epoch_0180.pth", map_location="cpu")
model.load_state_dict(state)

주의

  • 이 λͺ¨λΈμ€ λ…Όλ¬Έμš© truncation 강건성 νŠΈλž™ μ‚°μΆœλ¬Ό. challenge(과제용) νŠΈλž™μ˜ pallet-dope-challenge0123-ft-manual / pallet-dope-challengenight μ™€λŠ” λ³„κ°œ.
  • YOLO backend 와 혼용 μ‹œ R_fix=diag(-1,1,-1) 보정 ν•„μš”(+Z λΆ€ν˜Έ λ°˜λŒ€). DOPE 단독은 λΆˆν•„μš”.
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support