EfficientDet-D0: Compact Object Detector

• EfficientDet-D0 is a compact object detector that employs compound scaling to balance model size and detection accuracy.
• It integrates an EfficientNet-B0 backbone with a 3-layer BiFPN architecture to fuse multi-scale features using weighted sums and depthwise separable convolutions.
• The model attains 34.6 AP on COCO with only 3.9 million parameters and 2.5 billion FLOPs, demonstrating its efficiency for resource-constrained applications.

EfficientDet-D0 is the foundational variant of the EfficientDet family of object detectors, designed for optimal trade-off between model compactness and detection accuracy. It employs a compound scaling approach adapted from EfficientNet, a bi-directional feature pyramid network (BiFPN) architecture for multi-scale feature fusion, and lightweight prediction heads, resulting in a model that achieves strong detection performance with minimal computational overhead. EfficientDet-D0 processes $512 \times 512$ inputs, utilizes an EfficientNet-B0 backbone, a 3-layer 64-channel BiFPN, and achieves 34.6 AP on the COCO test-dev benchmark with only 3.9 million parameters and 2.5 billion multiply-adds (Tan et al., 2019).

1. Compound Scaling Approach

EfficientDet-D0 is configured using a compound scaling strategy that jointly scales three network dimensions—resolution ($r$), depth ($d$), and width ($w$)—in a principled manner. The scaling is controlled by a compound coefficient $\phi$, such that: $$r = \alpha^\phi,\quad d = \beta^\phi,\quad w = \gamma^\phi$$ with the empirical constraint $\alpha \cdot \beta^2 \cdot \gamma^2 \approx 2$, ensuring that increments in $\phi$ double computational cost. For object detection, this compound scheme defines:

• Backbone scaling via EfficientNet-B$\phi$ multipliers
• BiFPN width and depth: $W_{\rm BiFPN} = 64 \times (1.35)^\phi$, $r$0
• Box/class head depth: $r$1, width tied to BiFPN width
• Input resolution: $r$2

For EfficientDet-D0 $r$3, this configuration yields:

Parameter	Value for D0
$r$4	64
$r$5	3
$r$6	3
$r$7	512

This scaling design is inherited directly from EfficientNet and tailored for the needs of multi-scale object detection (Tan et al., 2019).

2. Input Resolution and Backbone Network

EfficientDet-D0 processes input images of size $r$8 using an EfficientNet-B0 backbone. EfficientNet-B0 itself is structured as follows:

• Stem: $r$9 conv, 32 channels, stride 2
• MBConv1, $d$0, 1 block, 16 channels (stride 1)
• MBConv6, $d$1, 2 blocks, 24 channels (stride 2)
• MBConv6, $d$2, 2 blocks, 40 channels (stride 2)
• MBConv6, $d$3, 3 blocks, 80 channels (stride 2)
• MBConv6, $d$4, 3 blocks, 112 channels (stride 1)
• MBConv6, $d$5, 4 blocks, 192 channels (stride 2)
• MBConv6, $d$6, 1 block, 320 channels (stride 1)

Feature maps are extracted at pyramid levels $d$7~$d$8; $d$9 and $w$0 are derived by successive $w$1 stride-2 depthwise convolutions on $w$2 and $w$3, respectively. The respective feature map sizes with 512×512 input are:

Level	Spatial Size
$w$4	$w$5
$w$6	$w$7
$w$8	$w$9
$\phi$0	$\phi$1
$\phi$2	$\phi$3

These five feature maps are passed to the BiFPN for subsequent fusion and processing (Tan et al., 2019).

3. BiFPN Architecture for Multi-Scale Fusion

The BiFPN (Bi-directional Feature Pyramid Network) within EfficientDet-D0 comprises $\phi$4 layers, with each layer performing multi-level bidirectional fusion. Each BiFPN node merges inputs via a learned, weighted sum: $\phi$5 where $\phi$6 are non-negative trainable coefficients and $\phi$7. All input feature maps are rescaled to uniform resolution prior to fusion, employing nearest-neighbor up/downsampling. The BiFPN exploits $\phi$8 depthwise separable convolutions at every spatial operation, followed by BatchNorm and the SiLU activation.

All BiFPN layers operate on 64 channels in D0. The design supports efficient, trainable exploitation of both top-down and bottom-up pathways at minimal additional computation, enabling effective multi-scale representation learning (Tan et al., 2019).

4. Prediction Heads and Anchor Design

On top of the BiFPN output, EfficientDet-D0 includes parallel box and class prediction heads, each structured as $\phi$9 consecutive $$r = \alpha^\phi,\quad d = \beta^\phi,\quad w = \gamma^\phi$$0 depthwise-separable convolutions (with BN and SiLU), using 64 channels. Both heads share parameters spatially across levels $$r = \alpha^\phi,\quad d = \beta^\phi,\quad w = \gamma^\phi$$1 to $$r = \alpha^\phi,\quad d = \beta^\phi,\quad w = \gamma^\phi$$2. The prediction heads process every BiFPN output to estimate bounding box coordinates (4 values per anchor) or class logits ($$r = \alpha^\phi,\quad d = \beta^\phi,\quad w = \gamma^\phi$$3 scores per anchor, $$r = \alpha^\phi,\quad d = \beta^\phi,\quad w = \gamma^\phi$$4 = number of classes).

Anchoring per feature level is realized as 9 anchors (3 aspect ratios × 3 scales), resulting in 45 anchors per spatial location when aggregated over five levels. EfficientDet-D0 applies identical head architecture for both regression and classification, promoting parameter efficiency without loss of task specialization (Tan et al., 2019).

5. Model Size, Efficiency, and Detection Performance

EfficientDet-D0 prioritizes compact model size and reduced arithmetic requirements. The summary of major metrics:

Metric	Value
Number of parameters	3.9 million
Multiply-Adds (FLOPs)	2.5 billion
COCO test-dev AP	34.6
COCO test-dev AP$$r = \alpha^\phi,\quad d = \beta^\phi,\quad w = \gamma^\phi$$5	53.0
COCO test-dev AP$$r = \alpha^\phi,\quad d = \beta^\phi,\quad w = \gamma^\phi$$6	37.1
COCO val split mAP	34.3

These performance results are achieved using a single model and single-scale evaluation, with no test-time augmentation other than Soft-NMS. The efficiency metrics highlight the emphasis on resource-constrained detection, making D0 the baseline configuration in the EfficientDet family (Tan et al., 2019).

6. Training Procedure and Implementation

EfficientDet-D0 is trained for 300 epochs using a global batch size of 128 on 32-core TPU v3 devices. Key hyperparameters and techniques include:

• Optimizer: SGD with 0.9 momentum, weight decay $$r = \alpha^\phi,\quad d = \beta^\phi,\quad w = \gamma^\phi$$7
• Learning rate: Linear warmup (0 $$r = \alpha^\phi,\quad d = \beta^\phi,\quad w = \gamma^\phi$$8 0.16 over 1 epoch), then cosine decay to zero
• BatchNorm: Synchronized across replicas, decay 0.99, $$r = \alpha^\phi,\quad d = \beta^\phi,\quad w = \gamma^\phi$$9
• Activation: SiLU (Swish)
• Losses: Focal loss ($\alpha \cdot \beta^2 \cdot \gamma^2 \approx 2$0, $\alpha \cdot \beta^2 \cdot \gamma^2 \approx 2$1) for classification; smooth-L1 for box regression
• Anchor configuration: 3 scales × 3 aspect ratios per level
• Data augmentation: Random horizontal flip and “scale jitter” in $\alpha \cdot \beta^2 \cdot \gamma^2 \approx 2$2 (input rescaled randomly by a factor in this range, then cropped/downsampled to $\alpha \cdot \beta^2 \cdot \gamma^2 \approx 2$3)
• Test-time: Only Soft-NMS is used for post-processing

These training settings ensure reproducibility of the reported metrics and robust convergence of the EfficientDet-D0 detector (Tan et al., 2019).