Modelpublic

cmes-deepvision/ACR-cuboid-v1.1.0

sourceHugging Faceupdated 3mo agoView on Hugging Face

0likes4downloads

Model Card

ACR Cuboid v1.1.0

3D cuboid detection for carton boxes on AGV — YOPO (You Only Predict Once).

transformers: 5.7.0

Changes from v1.0.x

—Combined dataset: 260429 (1,296 samples) + 260702 (678 samples) = 1,974 total
—260702 data: new carton box images with keypoint→cuboid conversion
—Base model: v1.0.3 (v7 best, checkpoint-1380)
—Training: 30 epochs, 7 GPU DDP, paper augmentation (HSV + translate + hflip)
—Loss coefficients: rot=5, center_2d=5, bbox=5, giou=2, depth=50, scale=50
—Backbone frozen, lr=1e-4 → 1e-5 (milestone 0.9167), maxgradnorm=0.1

Class prompts

Task	Prompts
carton_buffer	`item in box`, `top box`

Training results

—Final loss: 5.25 (90% reduction from initial)
—30 epochs, ~5h training time
—No overfitting observed (seen/unseen performance comparable)