Machine-Learning

Data Augmentation For Object Detection

Posted by Rico's Nerd Cluster on April 30, 2026

Training

Label a small but diverse calipers dataset.
Fine-tune RF-DETR Large.
Run it on new/unlabeled videos.
Find where it fails.
Add those failed frames back into the training set.
Retrain

False negative: Calipers are visible, but RF-DETR does not detect them. Add this frame to training with a correct calipers box.
False positive No calipers are present, but the model says something is calipers. Add this frame as a negative frame: image with zero boxes.

Tools: (5000 images)

Blender / BlenderProc: Very common for research.
NVIDIA Isaac Sim / Omniverse
a simpler OpenGL / pyrender style pipeline can work.
- Render the CAD over random backgrounds. object render + real background image compositing

Then you have: http://real_train/real_val/real_test/

Data Augmentation

Randomize:

matte grayblack plasticmetallic grayslightly rough surfaceslightly glossy surfacedifferent brightnessdifferent background color

Fix:
- camera distance
  - 50-60%: full object visible, centered-ish
  - 20-30%: object smaller, with table/background10-20%: partial crop / occlusion / close-up object occupies 20-70% of image width
- object scale in image
- full vs cropped views
- background/tablebbox correctness

get segmentation:

writer.initialize(
    output_dir=OUTPUT_DIR,
    rgb=True,
    bounding_box_2d_tight=True,
    semantic_segmentation=True,
)