How do I set tile overlap for small object detection in large images?

Key facts

What you need

Large images and a detection model trained on tiles of them

What you get

Small objects found across the whole image, each counted once

Image sources

Drone, satellite, microscope and high-resolution camera images

Runs on

Cloud API, or an edge device or server running Docker

How to set tile overlap for small objects, step by step

Six steps from a large image to every small object found once.

1. 01

   Connect Roboflow to your coding agent, ChatGPT or Codex

   Add the Roboflow MCP server to Claude, Cursor or Gemini CLI, or install the Roboflow plugin in ChatGPT or Codex. Your assistant can then upload and auto-label images, start training, build the Workflow and run the model from a prompt, at no extra cost beyond your Roboflow plan.

2. 02

   Measure your largest object in pixels

   Open a few images at full resolution and note the width, in pixels, of the largest object you need to find. That number sets your overlap.

3. 03

   Label your images and train on tiles of the same size

   Label your images in Roboflow Annotate, add tiling as a preprocessing step when you create a dataset version, and train on the tiles with Roboflow Train. Use the tile size you will run inference at.

4. 04

   Set the overlap to at least your largest object

   Divide your largest object's width by the tile width and round up. For 80 pixel objects on 640 pixel tiles that is 0.125, so use an overlap ratio of 0.15. Most cases land between 0.1 and 0.3. More overlap catches more objects on tile edges, but means more tiles and slower runs.

5. 05

   Run tiled inference and merge the duplicates

   Slice the image into overlapping tiles, run the model on each tile, and remove the duplicate detections where tiles overlap, so each object is counted once.

   PythonCopy

   import cv2
   import supervision as sv
   from inference import get_model
   
   model = get_model("your-project/1")
   image = cv2.imread("aerial_0031.jpg")
   LARGEST_OBJECT_PX = 80  # widest object you need to find
   
   def callback(tile):
       return sv.Detections.from_inference(model.infer(tile, confidence=0.4)[0])
   
   slicer = sv.InferenceSlicer(callback=callback, slice_wh=640, overlap_wh=LARGEST_OBJECT_PX + 20, overlap_filter="non_max_suppression")
   detections = slicer(image)
   print(len(detections), "objects found")
   # 214 objects found

   Rather not write code? Ask the Roboflow agent to set this up for you →

6. 06

   Build the tiling into a Workflow

   In a Workflow, chain the Image Slicer block with your overlap ratio, your model, and the Detections Stitch block set to drop duplicates (nms) or merge them (nmm). Run it with Roboflow Deploy in the cloud or on your own hardware.

Why 2 million engineers build on Roboflow

• Label faster

  Auto-label with SAM 3 or your own model, then fix the misses instead of drawing every box.

• RF-DETR

  Roboflow’s open-source real-time detection model, built for accuracy (opens in a new tab) on small objects.

• Logic built in

  Counting, zones, tracking and alerts are ready-made Workflow blocks, so there is less code to write.

• Your model, your weights

  Use your trained weights on every plan, including free, through the open-source Inference package.

• Run it anywhere

  Self-host on a workstation, Docker, an on-prem GPU or an edge device, or call the cloud API.