EfficientDet-D7: Scalable Object Detector

• EfficientDet-D7 is a state-of-the-art one-stage object detector that integrates EfficientNet-B6, a weighted BiFPN, and compound scaling to efficiently detect objects at multiple scales.
• The model's architecture leverages eight stacked BiFPN layers and specialized prediction heads, enabling robust feature fusion across scales with minimal computational overhead.
• Achieving 55.1% AP on COCO test-dev with 77 million parameters and 410 billion FLOPs, EfficientDet-D7 offers a favorable trade-off between accuracy and efficiency compared to previous detectors.

EfficientDet-D7 is a state-of-the-art one-stage object detector designed for efficient and accurate multi-scale object detection. Emerging from the EfficientDet family, it integrates a weighted bi-directional feature pyramid network (BiFPN), compound scaling, and an EfficientNet-B6 backbone for extracting image features. EfficientDet-D7 demonstrates a favorable trade-off between accuracy and computational cost, achieving 55.1% AP on COCO test-dev with a model containing 77 million parameters and 410 billion FLOPs when using single model and single scale, while being significantly smaller and faster than preceding systems (Tan et al., 2019).

1. Architecture Overview

EfficientDet-D7 comprises three principal components:

1. Backbone: EfficientNet-B6, pretrained on ImageNet, serves as the backbone and is responsible for encoding multi-scale features from the input image.
2. BiFPN: The Bidirectional Feature Pyramid Network is a sequence of identical layers that fuses features from backbone stages $P_3$ through $P_7$ using both top-down and bottom-up paths.
3. Class/Box Prediction Heads: Lightweight subnetworks are attached at each pyramid level and are responsible for classifying objects and regressing their bounding boxes.

This architectural composition realizes a high degree of parameter efficiency and enables robust multi-scale object representation.

2. BiFPN: Bidirectional Feature Pyramid Network

BiFPN is the critical component for cross-scale feature fusion. In EfficientDet-D7, five input features ($P_3$ to $P_7$) from the backbone, corresponding to scales 1/8, 1/16, 1/32, 1/64, and 1/128 of the $1536 \times 1536$ input, are utilized (for $P_3$, the resolution is $192 \times 192$; for $P_7$, $12 \times 12$).

Eight stacked BiFPN layers are employed. Each layer incorporates both top-down and bottom-up fusion, augmented by shortcut connections that maximize information flow while omitting nodes that aggregate a single input.

Feature fusion within BiFPN is realized via fast normalized fusion. For $N$ input features $P_7$0 at level $P_7$1, the output is

$P_7$2

where each $P_7$3 is a learnable weight (with non-negativity enforced via ReLU), and $P_7$4 prevents division by zero. This approach enables the network to learn the relative contribution of each input feature map with minimal overhead.

3. Compound Scaling Method

EfficientDet implements a compound scaling strategy to jointly scale network width, depth, and resolution using a global coefficient $P_7$5. Rather than disjointly tuning these dimensions, EfficientDet employs heuristic-based rules:

• Backbone: EfficientDet-D7 uses EfficientNet-B6, scaling width and depth following the EfficientNet series.
• BiFPN Width/Depth:
  • Width: $P_7$6
  • Depth: $P_7$7
• Prediction Head Depth: $P_7$8
• Input Resolution: $P_7$9

For EfficientDet-D7 ($P_3$0), this yields a $P_3$1 input, eight BiFPN layers of approximately 384 channels each, and five backbone stages from EfficientNet-B6.

4. Model Complexity and Empirical Results

EfficientDet-D7x, an extended version of D7 substituting EfficientNet-B7 as the backbone, demonstrates:

Model	Parameters (M)	FLOPs (B)	COCO test-dev AP (%)
D7x	77	410	55.1

Detailed COCO test-dev performance for D7x:

• $P_3$2: 74.3%
• $P_3$3: 59.9%
• $P_3$4
• $P_3$5
• $P_3$6

Relative to AmoebaNet + NAS-FPN (+AutoAugment), EfficientDet-D7x attains a $P_3$7 $P_3$8 improvement, with $P_3$9 fewer parameters (77M vs. 185M) and $P_7$0 fewer FLOPs (410B vs. 1317B).

5. Training and Implementation Details

Key training and implementation protocols essential for achieving EfficientDet-D7 performance are as follows:

• Optimization: SGD with momentum 0.9 and weight decay $P_7$1.
• Learning Rate: Linear warmup from 0 to 0.16 in the first epoch, then cosine decay to zero.
• Batch Size: 128, distributed across 32 TPUv3 cores; training is conducted for 600 epochs for D7/D7x.
• Data Augmentation: Horizontal flip, scale jittering in $P_7$2 range, followed by random cropping to $P_7$3.
• Loss Function: Focal loss (parameters: $P_7$4, $P_7$5), anchor aspect ratios $P_7$6.
• Normalization and Activation: Synchronized BatchNorm (decay 0.99, $P_7$7) after every convolution, SiLU (Swish-1) activations, Exponential Moving Average (EMA) of weights (decay 0.9998).
• Inference: Single-scale prediction with soft-NMS postprocessing.

6. Significance and Comparative Analysis

EfficientDet-D7 demonstrates that state-of-the-art object detection performance on COCO test-dev can be achieved with substantially reduced parameter count and arithmetic operations relative to prior detectors. Its architecture, combining EfficientNet-B6, depthwise-separable convolutions in BiFPN, weighted multi-scale fusion, and compound scaling, enables a new benchmark for single-model single-scale detection accuracy and complexity.

The model’s modular design permits adaptation to a broad range of deployment scenarios with varying computational constraints, as evidenced by scaling approaches that enable the entire EfficientDet family to fit diverse efficiency-accuracy trade-off envelopes (Tan et al., 2019). A plausible implication is that further advances in backbone architectures or feature fusion strategies might yield additional improvements under the same compound scaling regime.