pallet-dope-challengenight

pallet-dope-challengenight

DOPE (Deep Object Pose) 6D pose estimation model for forklift pallet detection, including night-time conditions. Fine-tuned on 14 real manual-GT capture sequences (8 day + 6 night) starting from challenge0123 (camera-facing v4 convention baseline).

Part of the Pallet 6D Pose Estimation project — geometry-aware self-training for automatic forklift alignment.

Note: this model was trained as challenge0123_ft_v2 (see header.txt's outf) and renamed to challengenight after training completed, to highlight night-time data inclusion.

Model overview

arch         : DOPE (VGG-19 backbone + belief maps + affinity fields)
keypoints    : 9  (8 cuboid corners + centroid, camera-facing v4 convention)
input        : 448x448 RGB
output       : belief maps (9ch) + affinity fields (16ch)
pallet dim   : 1.10 x 1.30 x 0.11 m  (KS T-11형)

Training

Weight       : final_net_epoch_0120.pth
Init weight  : challenge0123/final_net_epoch_0060.pth  (scratch -> 60 ep on mixed_v8 + chal_v1 + chal_v2)
Epochs       : 120  (60 -> 120, 60 ep fine-tune)
Batch size   : 8
LR           : 1e-4
Sigma        : 4.0 (belief Gaussian)
Image size   : 448
Workers      : 4
Seed         : 8055
Loss         : belief MSE + affinity (no symmetric / no geo / no struct / no rel)

Fine-tune data (14 manual-GT capture sequences)

Day (8 captures):

capturepallet02_manual_gt
capturepallet03_manual_gt
capturepallet04_manual_gt
capturepallet05_manual_gt
capturepallet07_manual_gt
capturepallet08_manual_gt
capturepallet09_manual_gt
capturepalletcad_manual_gt

Night (6 captures):

capturenight04_manual_gt
capturenight05_manual_gt
capturenight06_manual_gt
capturenight07_manual_gt
capturenight08_manual_gt
capturenight09_manual_gt

All captured with Intel RealSense D435i (640x480 @ fx=614.18, fy=614.31, cx=329.28, cy=234.53).

Files

final_net_epoch_0120.pth   trained model weight (final)
header.txt                 raw training Namespace + seed
README.md                  this file

Inference (depth_cam pipeline)

The matching pose-solving contract verified by twin_pnp_check.py (50/50 frame, reproj 2.89px, |dt|=0.085m):

PALLET_WIDTH_M  = 1.0    # mixed_v8_train label dim
PALLET_DEPTH_M  = 1.2
PALLET_HEIGHT_M = 0.15
PALLET_PNP_CONTRACT_Z180 = True   # Cuboid3d.vertices @ diag([-1,-1,+1])

task.yaml's nominal (1.1, 1.3, 0.11) is the spec value and does not match the actual label dim used in training.

Lineage

ImageNet pretrained VGG-19
        |
        v
challenge0123                 scratch, 60 ep, mixed_v8_train + chal_v1 + chal_v2
        |
        +--> challenge0123_ft_manual    20 ep on 6 day captures
        |
        +--> challengenight  <--- this model, 60 ep on 8 day + 6 night captures

Related

• GitHub: https://github.com/CanelE452/pallet-6d-pose
• Sibling model (day only): https://huggingface.co/CanelE452/pallet-dope-challenge0123-ft-manual
• Camera-facing v4 convention rationale: see project memory project_keypoint_convention_v4_conversion.md
• annotate.py PnP fix v4 (gravity invariant): see project_annotate_pnp_fix_v4_gravity.md

License

MIT