RF-DETR Fine-Tuning for a Custom Class
When fine-tuning RF-DETR on a custom object, the usual workflow is:
1
2
3
pretrained RF-DETR backbone/features
-> replace or reinitialize the classification head
-> train on the custom dataset
For example, if the target object is motor_sleeve, the new classifier head is trained for the custom label space:
1
2
0 motor_sleeve
background / no-object
It is not using the old COCO label space:
1
2
3
4
5
6
0 person
1 bicycle
...
67 cell phone
...
100 motor_sleeve
The old meaning of class 0 = person existed only in the original COCO classification head. During one-class fine-tuning, that head is resized or replaced, so class 0 now means motor_sleeve in the new checkpoint.
There is no numeric class-ID collision inside the fine-tuned checkpoint.
The pretrained model still keeps useful visual features in the backbone, such as edges, textures, shapes, and object-level representations. However, the final classification layer is retrained for the new dataset.
The main risk is not class-ID collision. The real risk is false-positive confusion. For example, the model may incorrectly detect visually similar objects as motor_sleeve if the custom dataset does not include enough negative examples or background variation.
If the goal is to detect both COCO classes and the new custom object, then the training setup must be different:
1
2
3
4
5
6
0 person
1 bicycle
...
67 cell phone
...
N motor_sleeve
In that case, the model needs a multi-class head and training data for both the original classes and the new class. Otherwise, the model may forget the old classes.
For the current one-class fine-tuning setup, the model should be treated as:
1
motor_sleeve detector + background/no-object
not as a COCO detector with one extra class.
