---

name: inference-visualization description: "Use MMDetection 3.3.0 public APIs for image, batch, folder, NumPy-array, video, large-image, visualization, prediction-dump, and deployment-route inference tasks." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

inference-visualization

Use this sub-skill when the task is to run MMDetection inference, save predictions or visualizations, process arrays/videos/large images, or choose a deployment path. For config/model-zoo selection route to configuration-model-zoo; for training/testing route to training-testing; for dataset conversion and metrics route to datasets-evaluation.

Fast Routes

• High-level image/folder/batch inference: use mmdet.apis.DetInferencer; see references/api-reference.md and references/workflows.md.
• Scripted image inference: run scripts/mmdet_infer_image.py for safe argparse around DetInferencer.
• Raw downstream objects: use return_datasamples=True or init_detector + inference_detector; see references/api-reference.md.
• Headless visualization/dumping: set out_dir, no_save_pred, and no_save_vis deliberately; see references/workflows.md.
• Video, webcam, and large images: prefer public API patterns and environment-aware caveats in references/workflows.md.
• TorchServe/MMDeploy/ONNX/TensorRT route choice: use references/deployment.md.
• Known failures: check references/troubleshooting.md before debugging package, device, GUI, checkpoint, or ABI errors.

Minimal Examples

from mmdet.apis import DetInferencer

inferencer = DetInferencer(model='rtmdet_tiny_8xb32-300e_coco', device='cpu')
result = inferencer('images/', batch_size=4, out_dir='outputs', no_save_pred=False, no_save_vis=True)

import mmcv
from mmdet.apis import init_detector, inference_detector

model = init_detector('config.py', 'checkpoint.pth', device='cpu')
image_bgr = mmcv.imread('image.jpg')
sample = inference_detector(model, image_bgr)

Packaged Helper

python scripts/mmdet_infer_image.py \
  image_or_folder rtmdet_tiny_8xb32-300e_coco --device cpu \
  --out-dir outputs --no-save-vis --save-pred

The helper uses public package APIs only and is safe for image files, image folders, model aliases, config files, local checkpoint files, and open-vocabulary prompt parameters supported by MMDetection.