Yolov11n Warehouse Pallets 640

YOLOv11 Warehouse Pallet Detector

A family of fine-tuned YOLOv11 models for real-time detection of pallets (wooden skid + stacked products) in warehouse environments. Available in 2 YOLOv11 sizes — nano (2.6M params, edge-ready) and small (9.4M params, best accuracy). Optimized for foreground pallet identification in operational warehouse settings with forklifts, racks, and dynamic lighting.

Try It

→ Interactive demo on Hugging Face Spaces — upload a warehouse image and run detection with any variant in the browser.

Examples

Real detections from the validation set using the nano @ 640p variant:

Clean foreground pallet

Clean pallet detection

Single foreground pallet detected at 0.86 confidence. Nano-640 is the edge workhorse — fast, clean, high-confidence on unambiguous shots.

Heavy motion blur

Motion blur pallet detection

Same camera in motion — pallet detected at 0.58 confidence despite the entire frame being smeared. Lower confidence than a clean shot, as expected, but the detection holds.

Cluttered multi-pallet scene

Multi-pallet detection

16 of ~20 ground-truth pallets detected with confidence ranging 0.30–0.83 in a cluttered aisle. Real deployments should expect this kind of variable-confidence output in dense scenes; post-filter by your target threshold.

Model Variants

The nano variant is published at four resolutions covering everything from microcontroller-class edge hardware to high-resolution production cameras. The small variant is published at 640p and 1280p, the two resolutions where additional capacity materially helps accuracy.

320p Models (Ultra-Edge)

Model	Params	Size (MB)	Resolution	mAP@0.5	mAP@0.5:0.95	Precision	Recall	GPU (ms)	CPU (ms)	Best For
YOLOv11n-320	2.6M	~5	320	TBD	TBD	TBD	TBD	TBD	TBD	Ultra-edge: ESP32-S3, Coral Edge TPU, RPi Zero, hobby/embedded

640p Models (Standard Resolution)

Model	Params	Size (MB)	Resolution	mAP@0.5	mAP@0.5:0.95	Precision	Recall	GPU (ms)	CPU (ms)	Best For
YOLOv11n-640	2.6M	~5	640	0.674	0.571	0.658	0.614	~5	~25	Edge / real-time on low-power devices
YOLOv11s-640	9.4M	~19	640	0.592	0.485	0.599	0.574	~7	~45	Balanced speed/accuracy

960p Models (Production Middle)

Model	Params	Size (MB)	Resolution	mAP@0.5	mAP@0.5:0.95	Precision	Recall	GPU (ms)	CPU (ms)	Best For
YOLOv11n-960	2.6M	~5	960	TBD	TBD	TBD	TBD	TBD	TBD	Jetson Orin Nano, 1080p cameras downsampled — better small-object recall than 640 without 1280 latency

1280p Models (Native High Resolution)

Model	Params	Size (MB)	Resolution	mAP@0.5	mAP@0.5:0.95	Precision	Recall	GPU (ms)	CPU (ms)	Best For
YOLOv11n-1280	2.6M	~5	1280	0.666	0.565	0.645	0.610	~18	~95	Edge with high-res cameras
YOLOv11s-1280	9.4M	~19	1280	0.680	0.580	0.646	0.628	~25	~170	Balanced, small pallet detection

All variants share the same training data, augmentation pipeline, and hyperparameters. Medium, large, and extra-large variants were tested but showed no accuracy improvement over small with the current dataset, so only nano and small are published. Metrics for the 320p and 960p nano variants are marked TBD — training is in progress.

Model Description

This model detects complete pallet units (wooden skid base + all products stacked on top) in warehouse imagery. It was trained on real-world warehouse photos captured during normal operations, making it robust to common warehouse conditions: motion blur, variable lighting, partial occlusions by forklifts and personnel, and cluttered backgrounds.

Unlike generic object detection models, this model is specifically trained to:

• Detect foreground pallets that are fully within the frame
• Distinguish pallets from visually similar structures (ceiling rafters, doors, rack uprights)
• Handle pallets stacked 2-high as separate detections per level
• Work reliably with motion-blurred images from moving cameras

Intended Use

• Warehouse automation: Real-time pallet counting and position tracking
• Forklift guidance: Detecting pallets in the robot/forklift field of view
• Inventory management: Automated pallet inventory from security or mounted cameras
• 3D warehouse mapping: Input to multi-view reconstruction pipelines for spatial pallet localization

Out of Scope

• Empty pallet (wooden skid only) detection without products
• Pallet type classification (EUR, GMA, block, stringer)
• Damaged pallet assessment
• Outdoor or non-warehouse environments

Training Details

Architecture

Variant	Base Model	Parameters	Input Resolutions	Classes	Framework
Nano	yolo11n.pt	~2.6M	320, 640, 960, 1280	1 (pallet)	Ultralytics 8.x
Small	yolo11s.pt	~9.4M	640, 1280	1 (pallet)	Ultralytics 8.x

The nano variant is trained at four resolutions (320, 640, 960, 1280) and the small variant at two (640, 1280), yielding 6 models total. All variants share the same architecture per size; only the training input resolution differs. Higher resolutions improve detection of small/distant pallets at the cost of slower inference.

Dataset

Split	Ratio	Description
Train	80%	Labeled warehouse images
Validation	15%	Held out for epoch-level evaluation
Test	5%	Held out for final evaluation

Training uses the full available dataset (no image cap). Exact counts depend on the number of labeled images at training time.

Labeling Pipeline: Images were auto-labeled using Qwen3.5-9B (natively multimodal vision-language model) with structured prompts to identify pallet bounding boxes, followed by human review of preview images. Negative examples (images with no pallets) are included as hard negatives.

Label Format: YOLO format (class_id, x_center, y_center, width, height) normalized 0-1.

Training Configuration

Hyperparameter	Value (640p)	Value (1280p)
Epochs	100 (early stopping, patience=20)	100 (early stopping, patience=20)
Batch Size	16	4 (auto-scaled)
Image Size	640	1280
Optimizer	SGD (Ultralytics default)	SGD (Ultralytics default)
Learning Rate	Auto-scaled	Auto-scaled
Augmentation	HSV jitter, horizontal flip, mosaic, scale	Same
Device	NVIDIA GPU (CUDA)	NVIDIA GPU (CUDA)

Data Augmentation Details:

hsv_h=0.015, hsv_s=0.7, hsv_v=0.4
degrees=10.0, translate=0.1, scale=0.5
fliplr=0.5, mosaic=1.0

Evaluation Results

Test Set Performance (All Variants)

Variant	mAP@0.5	mAP@0.5:0.95	Precision	Recall	GPU (ms)	CPU (ms)
Nano (n)	0.674	0.571	0.658	0.614	~5	~25
Small (s)	0.592	0.485	0.599	0.574	~7	~45

Benchmark Comparison

There is no established standard benchmark for warehouse pallet detection. The table below compares against results reported in published literature on similar (but not identical) datasets, to provide context for this model's performance.

Model	Dataset	Images	mAP@0.5	Precision	Recall	Source
This model (YOLOv11n)	EDITools Warehouse	Full dataset	0.572	0.600	0.563	This work
This model (YOLOv11s)	EDITools Warehouse	Full dataset	0.592	0.599	0.574	This work
NVIDIA SDG Pallet	Omniverse synthetic	~25,000	-	-	-	NVIDIA SDG Pallet Model
YOLOv8 (synthetic)	Unity synthetic	Synthetic	0.995	-	-	Pallet Detection From Synthetic Data (2025)
YOLOv8 (synthetic boost)	Synthetic + real	Custom	+69% stacked	-	-	Improving Pallet Detection Using Synthetic Data (2024)
YOLOv8	Custom warehouse	Custom	0.950	-	-	Semi-Autonomous Forklift (2025)
YOLOv11	Custom warehouse	Custom	-	0.93+	-	Semi-Autonomous Forklift (2025)
AM-Mask R-CNN	Complex warehouse	Custom	-	-	-	Enhanced Pallet Detection (2025)
Faster R-CNN	Industrial warehouse	1,344	0.89	-	-	IEEE Comparison (2020)
SSD	Industrial warehouse	1,344	0.85	-	-	IEEE Comparison (2020)
YOLOv4	Industrial warehouse	1,344	0.82	-	-	IEEE Comparison (2020)
YOLOv5 + ArUco	Custom + fiducial	Custom	-	0.995	-	Pallet Detection with YOLO + Fiducial (2023)
YOLOv8 + CBAM	Warehouse tracking	Custom	-	-	-	CBAM Pallet Tracking (2025)
YOLOX	Industrial	Custom	-	-	-	Digital Camera Pallet Detection

Notes on comparison:

• No standard benchmark exists for warehouse pallet detection (unlike COCO or KITTI for general/driving OD). Each study uses its own private dataset, making direct comparison difficult.
• The NVIDIA SDG model is the most production-ready alternative, trained on ~25K synthetic images via Omniverse, targeting pallet side-face centers/corners. It detects wood, metal, and plastic pallets but focuses on pallet pocket localization for forklift docking rather...