ACR-cuboid-v1.1.1

Single-class (carton box) 3D cuboid pose estimation model for ACR robotic picking.

Changes from v1.1.0

• Bugfix: Fixed 3 geometric augmentation bugs in rgbd_transforms.py:
  1. RandomRotation90RGBD: k=1 / k=3 center_2d and box label formulas were swapped
  2. RandomRotateRGBD: box width/height not updated after rotation
  3. MultiCropRGBD: stale labels returned when no objects survive crop
• Retrained 30 epochs with full geometric augmentation pipeline (v1.1.0 used paper-only augmentation)

Eval (checkpoint-2100, epoch 30)

Metric	Val	Test
3D IoU AP@0.50	99.97%	100.0%
3D IoU AP@0.75	94.85%	93.46%
COCO 2D AP	99.76%	99.45%
F1 (score≥0.5)	99.65%	99.83%
Translation (median)	10.28mm	10.33mm
Rotation (median)	1.41°	1.24°
Size LWH (median)	12.53mm	12.34mm

Architecture

• OpenVocabCuboidMapAnythingDINOFor3DObjectDetection
• Backbone: DINOv2-S (384-dim, 12 layers, registers)
• Decoder: 9-layer deformable DETR
• Parameters: 122.4M
• Input: 1280×1280

Compatibility

Drop-in replacement for v1.1.0 — identical config, architecture, and weight keys.