Object Detection

Overview

Object Detection node detects objects in an image using a custom-loaded model and a configurable post-processing pipeline.

This node gives full control over the detection stack: model selection, confidence filtering, NMS, label mapping, frame tiling, and overlay styling. Use this node when General Object Detection does not offer enough control, or when you need to run a domain-specific model.

Input

Input Image

image required

The image frame to analyze. Connect this to a camera or upstream image output.

Model path

string required

Path to the model directory to load for inference. See Model Directory Path for details.

Confidence threshold

number required advanced

Minimum confidence score to keep a detection. See Confidence Threshold for tuning guidance.

Range: 0.0 to 1.0
Default: 0.3

Overlap threshold

number required advanced

NMS overlap threshold (higher allows more overlap). See Overlap threshold for tuning guidance.

Range: 0.0 to 1.0
Default: 0.5

Class-agnostic NMS

boolean advanced

When enabled, NMS is applied across all classes. When disabled, NMS runs per class and you can configure NMS Class Group mappings (Class Label + Group ID).

Default: true

Use TensorRT if available

boolean advanced

When enabled, try to use TensorRT backend for inference (Nvidia GPU only). This may improve inference speed. For the first run, it will take some time to convert the model; subsequent runs will be much faster. If the model is not compatible, it will fallback to non-TensorRT inference with a warning.

Default: false

Divide Frame

boolean advanced

Splits a frame into smaller tiles according to the neural network input size, then merges results. Useful for improving recall on small objects in high-resolution images.

Default: false

When enabled, related advanced options appear:

• Overlap ratio (default 0.2) — overlap between divided tiles so edge objects are not missed.
• Greedy merge (default true) — merge detections from different tiles by overlap.
• Class agnostic merge (default false) — merge regardless of class vs same-class only.
• Use IoS for merge (default true) — use Intersection over Smaller instead of IoU.
• Merge match threshold (default 0.3) — overlap threshold for merge.
• Run standard prediction (default false) — also run full-frame prediction (helps large objects; slower).

Transform Contains

array advanced

Rule-based relabeling by text containment on detected labels.

Transform Labels

array advanced

Direct mapping from source labels to target labels.

Keep Labels

array advanced

Allow-list of labels to keep in final output.

Ignore Labels

array advanced

Deny-list of labels to remove from final output.

Overlay Results

boolean required advanced

Whether to draw detection overlays on output frame. See Overlay Results.

Default: true

Label colors

array advanced

Override default overlay colors for specific labels. Each entry has a Label and a Color (hex).

Custom Model Configs

boolean advanced

Reveals extra framework and architecture fields for non-.nam models. See Custom Model Configs.

Model Color Format

Color format of the image data sent to the model. If left blank, the default for the chosen architecture is used. Adjust if your model was trained with a different color format.

• RGB - Red, Green, Blue channel order.
• BGR - Blue, Green, Red channel order.

Scale

Multiplicative factor for the R, G, and B channels of the image before sending to the neural network. If left blank, the default for the chosen architecture is used. Used to normalize pixel values to the range expected by the model.

Mean

Mean values for the R, G, and B channels that are subtracted from the image before sending to the neural network. If left blank, the default for the chosen architecture is used. Used to normalize the image to match the data distribution the model was trained on.

Input Format

Tensor format for the image data sent to the neural network. If left blank, the default for the chosen architecture is used.

• NHWC - Batch x Height x Width x Channel (channels last).
• NCHW - Batch x Channel x Height x Width (channels first).

Input Type

Data type of the image input sent to the neural network.

• Uint8 - Pixel values are integers in the range 0-255.
• Float32 - Pixel values are 32-bit floating point numbers, often in the range 0.0-1.0 or normalized according to the model's training.

Post Process Type

Method for processing the raw output from the neural network before producing final detections. If left blank, the default for the chosen architecture is used.

• objdet_ontf2: ONNX TensorFlow 2

  "detection_classes": [N], [1_class_id, ..., n_class_id],
  "detection_scores": [N], [1_score, ..., n_score],
  "detection_boxes": [N*4], [1_y1, 1_x1, 1_y2, 1_x2, ..., n_y1, n_x1, n_y2, n_x2]

• objdet_onmmyolo: ONNX YOLO

  "labels": [N], [1_class_id, ..., n_class_id],
  "scores": [N], [1_score, ..., n_score],
  "boxes": [N*4], [1_x1, 1_y1, 1_x2, 1_y2, ..., n_x1, n_y1, n_x2, n_y2]

• objdet_onyoloworldv2: ONNX YOLO WorldV2

  "scores": [N*C], [1_class_a_score, 1_class_b_score, ..., n_class_a_score, n_class_b_score, ...],
  "boxes": [N*4], [1_x1, 1_y1, 1_x2, 1_y2, ..., n_x1, n_y1, n_x2, n_y2]

• objdet_onmmdet: ONNX MMDetection

  "dets": [N*5] e.g. [1_x1, 1_y1, 1_x2, 1_y2, 1_score, ..., n_x1, n_y1, n_x2, n_y2, n_score],
  "labels": [N] e.g. [1_class_id, ..., n_class_id]

Output

Overlay Image

image

Frame with detection overlays.

Detected Count

integer

Number of detected objects.

Detected Objects

array

Array of detected objects, each containing:

• bbox array: [x, y, width, height]
• label string
• confidence number

Tips

• Use a higher confidence threshold to reduce false positives.
• Tune Overlap threshold based on how close objects appear in your scene.
• When objects are small in high-resolution frames, frame division can improve detection recall.