module

EfficientDet

11 members

lucid.models.vision.efficientdet

EfficientDet — scalable compound-scaled detector (Tan et al., CVPR 2020).

Paper: "EfficientDet: Scalable and Efficient Object Detection"

Object Detection

Paper

Tan, Mingxing, et al. "EfficientDet: Scalable and Efficient Object Detection." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 10781–10790.

Overview

EfficientDet jointly applies the compound scaling idea of EfficientNet to the entire detection pipeline — backbone, feature network, and prediction heads — instead of scaling any single axis (width, depth, resolution) in isolation. A single compound coefficient $\varphi$ controls all three:

D_{\mathrm{bifpn}} = 3 + \varphi, \quad W_{\mathrm{bifpn}} = 64 \cdot 1.35^{\varphi}, \quad D_{\mathrm{head}} = 3 + \lfloor \varphi / 3 \rfloor, \quad R_{\mathrm{input}} = 512 + 128 \varphi,

so EfficientDet-D0 through D7 form a family with predictable accuracy–latency trade-offs without manual hyper-tuning.

The feature network is a BiFPN (Bi-directional Feature Pyramid Network). Unlike a one-way FPN, BiFPN adds bottom-up paths back into the pyramid and learns fast normalised fusion weights so each merged level is a convex combination of its inputs:

O = \sum_{i} \frac{w_i}{\varepsilon + \sum_j w_j} \cdot I_i, \qquad w_i \ge 0,

where the ReLU on $w_i$ (here written as a non-negativity constraint) keeps fusion stable without an explicit softmax, and the small $\varepsilon = 10^{-4}$ prevents division by zero. BiFPN is then stacked $D_{\mathrm{bifpn}}$ times.

The detection heads (classification + box) share weights across levels and use sigmoid + focal loss for classification, the same anchor-based formulation as RetinaNet. Combined with the EfficientNet-B\ $_{\varphi}$ backbone, the result is the Pareto-optimal one-stage detector family of its time — EfficientDet-D7 reaches 55.1 AP on COCO at far fewer FLOPs than contemporaries.

Classes

dataEfficientDetConfig1 methods

Configuration for EfficientDet.

classEfficientDetForObjectDetection3 methods

EfficientDet object detector (Tan et al., CVPR 2020).

Functions

fnefficientdet_config→ EfficientDetConfig

Build the canonical EfficientDet-D{phi} config from the compound table.

fnefficientdet_d0→ EfficientDetForObjectDetection

EfficientDet-D0 ( $\varphi = 0$ ).

fnefficientdet_d1→ EfficientDetForObjectDetection

EfficientDet-D1 ( $\varphi = 1$ ).

fnefficientdet_d2→ EfficientDetForObjectDetection

EfficientDet-D2 ( $\varphi = 2$ ).

fnefficientdet_d3→ EfficientDetForObjectDetection

EfficientDet-D3 ( $\varphi = 3$ ).

fnefficientdet_d4→ EfficientDetForObjectDetection

EfficientDet-D4 ( $\varphi = 4$ ).

fnefficientdet_d5→ EfficientDetForObjectDetection

EfficientDet-D5 ( $\varphi = 5$ ).

fnefficientdet_d6→ EfficientDetForObjectDetection

EfficientDet-D6 ( $\varphi = 6$ ).

fnefficientdet_d7→ EfficientDetForObjectDetection

EfficientDet-D7 ( $\varphi = 7$ ).