d_fine_l_objects365-coco_hgnet_v2_b4_pp-imagenet22k

Model Card for d_fine_l_objects365-coco_hgnet_v2_b4_pp-imagenet22k

A D-FINE large object detector with an HGNet v2 B4 backbone, pretrained on Objects365-2020 and fine-tuned on COCO 2017. Training used multi-resolution inputs sampled from 480px to 800px.

An inference-optimized, structurally reparameterized checkpoint is also provided as d_fine_l_objects365-coco_hgnet_v2_b4_pp-imagenet22k_reparameterized.

Important: The reparameterized checkpoint fuses compatible convolution and normalization branches and removes training-only decoder components. It is intended for inference and deployment. For continued training or fine-tuning, use the standard checkpoint, d_fine_l_objects365-coco_hgnet_v2_b4_pp-imagenet22k, and reparameterize the model only after training is complete.

Custom Kernels: This model uses optimized custom kernel for Deformable Attention operations. If you encounter compilation issues or prefer to use pure PyTorch implementations, set the environment variable DISABLE_CUSTOM_KERNELS=1 before loading the model.

Model Details

• Model Type: Object detection

• Model Stats:

  • Params (M): 31.3
  • Input image size: 640 x 640

• Dataset: COCO 2017 (80 classes)

• Papers:

  • D-FINE: Redefine Regression Task in DETRs as Fine-grained Distribution Refinement: https://arxiv.org/abs/2410.13842

• Metrics:

  Input size	mAP	mAP (reparam.)	mAP@50	mAP@50 (reparam.)	mAP@75	mAP@75 (reparam.)
  512 x 512	54.00	53.99	71.85	71.84	58.64	58.64
  576 x 576	55.47	55.49	73.07	73.08	60.31	60.32
  608 x 608	55.77	55.77	73.34	73.37	60.81	60.79
  640 x 640	56.09	56.08	73.51	73.51	60.98	60.97
  672 x 672	56.21	56.21	73.56	73.56	61.21	61.21
  704 x 704	56.43	56.43	73.87	73.87	61.54	61.53

Model Usage

Object Detection

import birder
from birder.inference.detection import infer_image

# Option 1: manual setup (more control over preprocessing)
net, model_info = birder.load_pretrained_model("d_fine_l_objects365-coco_hgnet_v2_b4_pp-imagenet22k", inference=True)

# Get the image size the model was trained on
size = birder.get_size_from_signature(model_info.signature)

# Create an inference transform
transform = birder.detection_transform(size, model_info.rgb_stats, dynamic_size=model_info.signature["dynamic"])

# Option 2: helper (quick start with default preprocessing)
net, model_info, transform = birder.load_pretrained_model_and_transform("d_fine_l_objects365-coco_hgnet_v2_b4_pp-imagenet22k", inference=True)

image = "path/to/image.jpeg"  # or a PIL image, must be loaded in RGB format
detections = infer_image(net, image, transform)
# detections is a dict with keys: 'boxes', 'labels', 'scores'
# boxes: torch.Tensor with shape (N, 4) in [x1, y1, x2, y2] format
# labels: torch.Tensor with shape (N,) containing class indices
# scores: torch.Tensor with shape (N,) containing confidence scores

Reparameterized Inference

Use the _reparameterized checkpoint for inference-oriented deployment. Birder automatically constructs the matching reparameterized architecture, no additional conversion flag is needed.

import birder
from birder.inference.detection import infer_image

weights = "d_fine_l_objects365-coco_hgnet_v2_b4_pp-imagenet22k_reparameterized"
net, model_info, transform = birder.load_pretrained_model_and_transform(weights, inference=True)

image = "path/to/image.jpeg"  # or a PIL image, must be loaded in RGB format
detections = infer_image(net, image, transform)

Citation

@misc{peng2024dfineredefineregressiontask,
      title={D-FINE: Redefine Regression Task in DETRs as Fine-grained Distribution Refinement},
      author={Yansong Peng and Hebei Li and Peixi Wu and Yueyi Zhang and Xiaoyan Sun and Feng Wu},
      year={2024},
      eprint={2410.13842},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2410.13842},
}