YOLOv4→YOLOv11 Posture AI on Kria KV260

comparison of yolov4 and yolo v11
kria kv260_2
yolo v4 vs yolo v11
kria kv260_1
vitis ai
petalinux_os_build
vivado dpu block design

Inspiration

Modern detectors like YOLOv11 are strong in accuracy and size, but deploying them on edge DPU hardware is not drop-in. Ops, activations, and head design often break quantize/compile for Vitis AI + DPUCZDX8G.

We used a fixed vision task only as a workload, and focused on the real goal: make a newer YOLO family run efficiently on Kria KV260 and compare it to a known baseline.

What we optimized

Stage	Model	Focus
Baseline	YOLOv4-Leaky INT8	Measured reference on KV260 DPU
Optimization	YOLOv11n INT8	Surgery → quant → compile → DPU runtime

Same board (KV260), same DPU (DPUCZDX8G), same task framing—so differences come from model + deployment path, not from a new application story.

How we built the optimization path

Baseline lock-in — YOLOv4 already quantized and running on KV260 (~11.17 FPS, ~95 ms latency, strong INT8 accuracy). This is the reference, not the product pitch.
YOLOv11n train — Nano variant (~2.6M params) trained for the same 2-class head so the deploy path stays comparable.
Model surgery for DPU — Replaced unsupported or fragile patterns (e.g. SiLU → Hardsigmoid, safer splits, head simplified for tracing) so Vitis AI 3.5 could quantize and compile.
Quantize + compile — INT8 path to xmodel for DPUCZDX8G.
On-board runner — VART load, correct fix_point dequant, multi-scale DFL decode, NMS; verified end-to-end on hardware.

Results (optimization view)

YOLOv4 baseline (KV260)

End-to-end ~11.17 FPS, ~95 ms
INT8 accuracy held up well after quant

YOLOv11n optimized path (KV260)

DPU inference ~12.7 ms per frame
Full train → surgery → INT8 → xmodel → board path working
Detections on-device; confidence still softer than YOLOv4 after INT8 (expected next lever: calib/QAT)

So: baseline = proven full-stack INT8 deploy; optimization = newer, smaller detector brought onto the same DPU with explicit surgery and runtime fixes.

Challenges

Graph compatibility — Stock YOLOv11 did not quantize cleanly; surgery was mandatory.
Tracing / quant failures — Head and tensor layouts needed simplification before vai_q / compile succeeded.
Wrong post-process looks like model failure — DPU ran in ~12 ms with 0 boxes until dequant + DFL decode matched the 66-channel heads (4×16 + 2).
INT8 score drop — YOLOv11 runs; calibration quality still limits confidence vs YOLOv4.

What we learned

On Arm + DPU, optimization work is mostly deployment engineering: ops, quant, compile, and decode—not only training.
A measured baseline on the same board makes claims honest.
Model surgery is a first-class step for YOLO-class models on DPUCZDX8G.
Sub-13 ms DPU time is achievable for YOLOv11n; closing the accuracy gap after INT8 is the next optimization target.

What’s next

Stronger INT8 recovery (better calib set / QAT) for YOLOv11n
Fair board-level FPS comparison (full pre/post included), not only pure DPU time
Optional further graph trimming for DPU utilization

Bottom line: This project is about bringing YOLOv11n through Vitis AI surgery and INT8 onto KV260, using YOLOv4 as the on-hardware baseline—not about building a posture product narrative.

Built With

arm
dpuczdx8g
edge-ai
int8
kria-kv260
opencv
physical-ai
python
pytorch
vitis-ai
yolov11
yolov4

Submitted to

Arm Create: AI Optimization Challenge

Created by

Worked on INT8 quantization and DPU compilation for YOLOv11n (Vitis AI 3.5 → xmodel for DPUCZDX8G), model surgery for graph compatibility, and on-board validation of the optimized model on Kria KV260.

ROHAN CHERALA
AI Model Training & Development

Contributed to the development and optimization of the deep-learning models used for edge-AI inference on the AMD Kria KV260 platform. My primary responsibility focused on the AI model development pipeline, including dataset preparation, model training, evaluation, and performance analysis.

- Prepared and validated the vision dataset for the target two-class object-detection task, ensuring suitable data quality and class distribution for training.
- Trained and evaluated YOLO-based object-detection models, including the YOLOv11n nano architecture, with a focus on achieving a lightweight model suitable for resource-constrained edge hardware.
- Experimented with model configurations, training parameters, input resolution, and optimization strategies to balance detection accuracy, model complexity, and inference efficiency.
- Performed comparative evaluation against the existing YOLOv4-Leaky baseline, analyzing model accuracy and inference behavior to establish a meaningful performance reference.
- Investigated the model architecture from a deployment perspective and collaborated on identifying operations and network components that were incompatible or inefficient for the target DPU.
- Supported the model-surgery and quantization workflow, adapting the trained YOLOv11n model for INT8 inference and subsequent Vitis AI compilation.
- Analyzed post-quantization detection performance and helped identify the accuracy degradation introduced during INT8 conversion, establishing calibration and QAT as the next optimization direction.
- Contributed to validating the complete training → optimization → quantization → deployment pipeline, resulting in a YOLOv11n model successfully running on the Kria KV260 DPU with approximately 12.7 ms DPU inference time per frame.

Overall, my contribution bridged the gap between AI model training and hardware-aware deployment, ensuring that the trained model was not only accurate in software but also suitable for efficient execution on FPGA/DPU-based edge hardware.

Bandari Govardhan
Trained and optimized the YOLOv11 model for posture detection and prepared it for deployment on the Kria KV260 using Vitis AI.

Karthik Pakala
ENUMULA BHUVAN SHEKAR REDDY