“Object Detection Models: Balancing Speed, Accuracy and Efficiency,” a Presentation from Union.ai

For the full video of this presentation, please visit: https://www.edge-ai-vision.com/2025/10/object-detection-models-balancing-speed-accuracy-and-efficiency-a-presentation-from-union-ai/

Sage Elliott, AI Engineer at Union.ai, presents the “Object Detection Models: Balancing Speed, Accuracy and Efficiency,” tutorial at the May 2025 Embedded Vision Summit.

Deep learning has transformed many aspects of computer vision, including object detection, enabling accurate and efficient identification of objects in images and videos. However, choosing the right deep neural network-based object detector for your project, particularly when deploying on lightweight hardware, requires consideration of trade-offs between accuracy, speed and computational efficiency.

In this talk, Elliott introduces the fundamental types of DNN-based object detectors. He covers models such as Faster R-CNN for high-accuracy applications and single-stage models such as YOLO and SSD for faster processing. He discusses lightweight architectures, including MobileNet, EfficientDet and vision transformers, which optimize object detection for resource-constrained environments. You will learn the trade-offs between object detection models for your computer vision applications, enabling informed choices for optimal performance and deployment.

1. Object Detection Models:
Balancing Speed, Accuracy and
Efficiency
Sage Elliott
AI Engineer
Union.ai

2. Object Detection 101: The Basics
•What is object detection?
oDetects and classifies multiple objects in
an image or video by drawing bounding
boxes around them and assigning labels.
•What data is needed?
oA dataset of labeled images, where each
object is annotated with a bounding box
and a class label.
© 2025 union.ai | Sage Elliott 2

3. Why Object Detection Matters
Real-world use cases (sample):
oAutonomous Vehicles & Robotics
oSecurity
oMobile Apps
Real-world systems must balance:
oSpeed —fast enough for real-time use (if needed)
oAccuracy —reliable predictions under noise & variation
oEfficiency —low compute cost, especially on edge
devices
© 2025 union.ai | Sage Elliott 3

4. Evolution of Object Detection
•Traditional CV: Handcrafted features(edge
detectors, HOG) + classical classifiers (SVMs,
decision trees)
•Deep learning: Learns features and patterns
directly from raw image data using
convolutional neural networks (CNNs)
oTransformers:Recent models (DETR, DINO) use
attention mechanisms for end-to-end object
detection
© 2025 union.ai | Sage Elliott 4

5. Measuring Object Detection
Measuring object detection is more complex
than classification.
•Intersection over union (IoU)
•Average precision (AP) on a threshold
•Mean average precision (mAP) | .50 -.95
•mAPon Coco dataset a common metric
mAPmay seem low on models. But they can have
very high AP on higher thresholds.
© 2025 union.ai | Sage Elliott 5

6. Major Architectures for Deep Learning Object Detection
•Two-stage detectors (Faster R-CNN, DETR)
oHigh accuracy
oSlower, more compute-intensive
•One-stage detectors (YOLO, SSD)
oFast, real-time friendly
oHistorically less accurate (but improving)
© 2025 union.ai | Sage Elliott 6

7. Two-Stage Models: Overview
•Stage 1: Region proposal
oIdentify potential object locations (using
Region Proposal Network (RPN)
•Stage 2: Classification + box refinement
oPredict class label and refine bounding box
coordinates for each proposal
•High accuracy, especially in complex scenes
with small or overlapping objects
•Slower inference time, higher computational
cost
© 2025 union.ai | Sage Elliott 7
Non-Max Suppression (NMS)
With confidence threshold

8. Spotlight: RCNN / Faster R-CNN
•Uses a CNN backbone (ResNet) for
feature extraction
•Generates region proposals with an
RPN
•Classifies each region and refines
bounding box coordinate
•Higher-accuracy applications where
speed is less critical
© 2025 union.ai | Sage Elliott 8

9. One-Stage Models: Overview
•Detection and classification happen in one
forward pass
•Predict class labels and bounding boxes
directly from feature maps
•Designed for speed and efficiency, suitable
for real-time applications
•Typically faster, but may trade off some
accuracy in complex scenes
© 2025 union.ai | Sage Elliott 9

10. Spotlight: SSD (Single Shot Detector)
•Uses anchor boxes to detect objects and
class labels in a single forward pass
•Balances speed and accuracy—faster than
two-stage models like Faster R-CNN
•Well-suited for edge deployments where
performance and reliability matter
•Leverages multi-scale feature maps to
detect objects of varied sizes effectively
© 2025 union.ai | Sage Elliott 10
Default SSD has 8732 Boxes

11. Spotlight: YOLO (You Only Look Once)
•Predicts bounding boxes and class labels
in one forward pass
•Evolved from YOLOv1 to YOLOv9with
major improvements in speed and
accuracy
•Extremely fast and efficient—ideal for
real-time applications and edge devices
•Strong balance of speed and decent
accuracy across a wide range of tasks
© 2025 union.ai | Sage Elliott 11

12. Spotlight: DETR(DEtectionTRansformer)
•Uses a CNN backbone to extract image
features
•Applies a Transformer encoder-decoder to
model global relationships
•Predicts a fixed set of objects using learned
object queries
•Complex visual scenes or vision-language
tasks where global reasoning and a
simplified detection pipeline are beneficial
© 2025 union.ai | Sage Elliott 12
Attention-based model –we can
visualize what the network is
looking at to make predictions.
Images from DETR Paper

13. Classification "Backbones" for Object Detection
•Most object detection models use a
classification backbone to extract visual
features
•These can be swapped to optimize for
speed, accuracy, or efficiency
oMobileNet
oEfficientNet
oResNet50
oSwin Transformers
© 2025 union.ai | Sage Elliott 13

14. How to Choose the Right Model
•What’s your available hardware (CPU, GPU, edge device)?
•How important is accuracy vs. inference speed?
•Do you need real-time performance, or can you afford latency?
•Are you detecting simple objects or complex scenes?
•Experiment
© 2025 union.ai | Sage Elliott 14

15. Other Deployment Considerations
Beyond model choice, optimize for speed,
size, and deployment environment:
•Model quantization & pruning
•Formats: ONNX, TensorRT, LiteRT(TFlite)
•Serving frameworks: TorchServe,
OpenVINO, Ray, Union.ai
© 2025 union.ai | Sage Elliott 15
Visual (Library
Logos?)

16. Summary & Final Thoughts
•Trade-offs are everywhere—there's no one-size-fits-all model
•Know your constraints —orfigure them out fast
•Optimize for real-world use —not just benchmarks
•Build an efficient pipeline —make it easy to test and compare models quickly
© 2025 union.ai | Sage Elliott 16

17. R-CNN
https://arxiv.org/abs/1506.01497v3
DETR
https://arxiv.org/abs/2005.12872v3
YOLO
https://arxiv.org/abs/1506.02640
SSD
https://arxiv.org/abs/1512.02325
Get these links & examples:
https://github.com/sagecodes
© 2025 union.ai | Sage Elliott 17
Resources

18. Thank You and Stay Connected!
•Thank you!
•I help build efficient AI pipelines at
Union.ai (including computer vision)
•Check out Union.ai and RSVP for my
upcoming object detection workshop!
© 2025 union.ai | Sage Elliott 18