In this paper, we systematically study neural network architecture design choices for object detection and propose several key optimizations to improve efficiency. First, we propose a weighted bi-directional feature pyramid network (BiFPN), which allows easy and fast multiscale feature fusion. Second, we propose a compound scaling method that uniformly scales the resolution, depth, and width for all backbone, feature network, and box/class prediction networks at the same time. Based on these optimizations, we develop a new family of object detectors, called EfficientDet, which consistently achieve much better efficiency than prior art across a wide spectrum of resource constraints. In particular, EfficientDet-D7 achieves 55.1 AP on COCO test-dev with 77M parameters and 410B FLOPs, being 4x-9x smaller and using 13x-42x fewer FLOPs than previous detectors.

Object detection has been driven by the development of more powerful but also increasingly expensive backbone networks. While accuracy has steadily improved, the computational cost has grown even faster, making deployment on resource-constrained devices challenging. A natural question is: is it possible to build a scalable detection architecture that achieves both higher accuracy and better efficiency?

We identify two main challenges: (1) efficient multi-scale feature fusion — previous methods like FPN treat all input features equally, but "features at different resolutions often contribute to the fused output feature unequally"; (2) model scaling — prior works scale only the backbone or feature network independently, leading to suboptimal accuracy-efficiency trade-offs.

Feature Pyramid Network (FPN) introduces a top-down pathway to fuse multi-scale features. PANet adds a bottom-up path on top of FPN. NAS-FPN uses neural architecture search to find the optimal feature network topology, but it is irregular, hard to interpret, and expensive to modify. For model scaling, EfficientNet demonstrated that compound scaling of depth, width, and resolution is more effective than single-dimension scaling for image classification. We extend this idea to object detection, jointly scaling the backbone, BiFPN, and prediction networks.

We propose BiFPN (Bi-directional Feature Pyramid Network) with two key improvements. First, we remove nodes with only one input edge, since a node with only one input cannot fuse features and thus contributes less. This simplification leads to a more efficient bi-directional (top-down + bottom-up) topology. Second, unlike FPN which treats all inputs equally, we add learnable weights for each input during feature fusion, allowing the network to learn the importance of different resolution features. We use fast normalized fusion: w_i / (sum(w_j) + epsilon), which achieves comparable accuracy to softmax-based fusion but runs 30% faster on GPU.

Inspired by EfficientNet, we propose a compound scaling method for object detection that jointly scales up all dimensions of the detector. Given a compound coefficient phi, we scale the backbone network using EfficientNet-B(phi), the BiFPN width and depth proportionally, and the input resolution using a heuristic formula. The box/class prediction network depth also increases with phi. This produces a family from EfficientDet-D0 (3.9M params, 2.5B FLOPs) to EfficientDet-D7 (77M params, 410B FLOPs), covering a wide spectrum of accuracy-efficiency trade-offs with a single unified architecture.

On COCO test-dev, the EfficientDet family achieves consistently better accuracy-efficiency trade-offs compared to prior detectors. EfficientDet-D0 achieves 34.6 AP with only 3.9M parameters, matching YOLOv3's 33.0 AP while being 28x smaller. EfficientDet-D7 achieves 55.1 AP, surpassing the previous best NAS-FPN + AmoebaNet's 50.7 AP while using 4x fewer parameters and 13x fewer FLOPs. Compared to RetinaNet, EfficientDet-D1 achieves similar accuracy (40.2 vs. 40.4 AP) with 8x fewer parameters and 21x fewer FLOPs.

Ablation studies demonstrate the effectiveness of each component. BiFPN improves AP by 4.1 points over the standard FPN baseline. Weighted feature fusion adds 0.4 AP over unweighted BiFPN. The compound scaling consistently outperforms single-dimension scaling: at 50B FLOPs, compound scaling achieves 46.3 AP vs. 43.9 AP for scaling resolution alone. These results confirm that both BiFPN and compound scaling contribute significantly to the final performance.

In this paper, we systematically study various architecture design choices for efficient object detection and propose BiFPN for efficient multi-scale feature fusion and a compound scaling method to jointly scale up all components of the detector. These optimizations lead to a new family of detectors, EfficientDet, that achieves 4x-9x parameter reduction and 13x-42x FLOPs reduction while maintaining state-of-the-art accuracy.