how to add background images (images to ignore)

[cyberpunk-edgerunner]

I want my rf-detr model to treat some images as background
will it work if I put them inside the train folder without any annotations ??

[JESUSROYETH]

Hi @cyberpunk-edgerunner 👋

Yes, background images help — RF-DETR is a DETR-style model, so it learns a "no object" prediction for every query, and feeding it images with no objects teaches it to predict no boxes on that kind of scene and lowers false positives.

But just dropping the files into the train folder is not enough, and that part is easy to miss. RF-DETR does not scan the directory, its dataloader inherits from torchvision CocoDetection, which iterates over the images array of your _annotations.coco.json. Any file that is not registered there is silently ignored.

So the right way is to add the image to the images list in the COCO json and give it no matching entries in annotations. The loader then returns an empty target (0 boxes, 0 labels) and the image trains as a pure background sample:

// _annotations.coco.json
"images": [
    {"id": 1, "file_name": "obj.jpg",        "width": 64, "height": 64},
    {"id": 2, "file_name": "background.jpg",  "width": 64, "height": 64}   // no annotations -> background
],
"annotations": [
    {"id": 1, "image_id": 1, "category_id": 1, "bbox": [10, 10, 20, 20], "area": 400, "iscrowd": 0}
]

I checked this on rfdetr 1.8.3. I built a dataset with 3 image files on disk but registered only 2 in the json — len(dataset) came back as 2 (the orphan file was dropped), and the registered image with no annotations returned boxes=(0, 4), labels=[] with no crash ..

Versions: rfdetr 1.8.3, torch 2.12.1, torchvision 0.27.1, Python 3.12.3.

Hope this helps!

[cyberpunk-edgerunner]

That helps a lot And I was looking at the documentation; it states that the rf-detr also accepts annotations in YOLO format if that is the case, can I just leave the background images in the train/images and val/images folder without any annotations for those ones in train/labels and val/labels ??

…

On Sat, Jul 4, 2026 at 11:38 PM Jesús Royeth ***@***.***> wrote: *JESUSROYETH* left a comment (roboflow/rf-detr#1186) <#1186 (comment)> Hi @cyberpunk-edgerunner <https://github.com/cyberpunk-edgerunner> 👋 Yes, background images help — RF-DETR is a DETR-style model, so it learns a "no object" prediction for every query, and feeding it images with no objects teaches it to predict no boxes on that kind of scene and lowers false positives. But just dropping the files into the train folder is not enough, and that part is easy to miss. RF-DETR does not scan the directory, its dataloader inherits from torchvision CocoDetection, which iterates over the images array of your _annotations.coco.json. Any file that is not registered there is silently ignored. So the right way is to add the image to the images list in the COCO json and give it no matching entries in annotations. The loader then returns an empty target (0 boxes, 0 labels) and the image trains as a pure background sample: // _annotations.coco.json"images": [ {"id": 1, "file_name": "obj.jpg", "width": 64, "height": 64}, {"id": 2, "file_name": "background.jpg", "width": 64, "height": 64} // no annotations -> background ],"annotations": [ {"id": 1, "image_id": 1, "category_id": 1, "bbox": [10, 10, 20, 20], "area": 400, "iscrowd": 0} ] I checked this on rfdetr 1.8.3. I built a dataset with 3 image files on disk but registered only 2 in the json — len(dataset) came back as 2 (the orphan file was dropped), and the registered image with no annotations returned boxes=(0, 4), labels=[] with no crash .. Versions: rfdetr 1.8.3, torch 2.12.1, torchvision 0.27.1, Python 3.12.3. Hope this helps! — Reply to this email directly, view it on GitHub <#1186?email_source=notifications&email_token=BSIM7DS25E3JM7GAV7FRRCD5DFB2RA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBYGMZTIMRSHE32M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4883342297>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BSIM7DTVR7DPZQTLGMQUFHL5DFB2RAVCNFSNUABFKJSXA33TNF2G64TZHM4TKMJVGM2DENRQHNEXG43VMU5TIOBQHE4DMNJYGYYKC5QC> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/BSIM7DQSOD7DXCIXCLQQDS35DFB2RA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBYGMZTIMRSHE32M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android <https://github.com/notifications/mobile/android/BSIM7DSZUFNY235M4FB45OL5DFB2RA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBYGMZTIMRSHE32M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you were mentioned.Message ID: ***@***.***>

[JESUSROYETH]

Yes, you can. In YOLO format a background image is just an image with no matching .txt label file (an empty .txt works the same). Drop them in the train images folder without a label and rf-detr treats them as background samples with empty detections, so they push the model towards fewer false positives.

One thing to keep in mind: this used to be broken — images without a label were silently skipped — but it was fixed in #915, merged back in April. So as long as you are on a recent version (1.6 or newer, current is 1.8.3) you are fine. On anything older they just get dropped without warning ..

[JESUSROYETH]

Yes, for the YOLO format it works like you expected: you drop the background image into train/images (and valid/images) and you don't create any .txt for it in labels/. The loader then treats it as a background sample with 0 boxes.

The reason is simple: the two formats have opposite defaults, and that is easy to miss. In COCO an image that is not registered in the json gets silently ignored — that is why in my first answer you had to add it to the images list. In YOLO it is the other way around: RF-DETR scans train/images directly and, for every image, it looks for a matching labels/<name>.txt. If that file is missing it just returns an empty target instead of skipping the image, an empty .txt (0 lines) works the same way.

I checked this on rfdetr 1.8.3. I built a YOLO dataset with images in train/images, gave one a label file and left another with no .txt at all — len(dataset) counted both, and the unlabelled image returned boxes=(0, 4), labels=[] with no crash:

background.jpg   boxes=(0, 4)  labels=[]     # no .txt   -> background
empty_lbl.jpg    boxes=(0, 4)  labels=[]     # empty .txt -> background
obj.jpg          boxes=(1, 4)  labels=[0]

One thing to keep in mind: the dataset still has to be a valid YOLO layout (a data.yaml/data.yml, plus train/ and valid/, each with images/ and labels/), otherwise RF-DETR won't detect it as YOLO in the first place.

Versions: rfdetr 1.8.3, torch 2.12.1, torchvision 0.27.1, Python 3.12.3.

Hope this helps!

[cyberpunk-edgerunner]

Thanks for your detailed explanation. I have another question: 1. Why is RFDetr using a different number of positional encodings than DINOv2? Can you tell me the reason behind this? I used both large and 2XLarge, and in both cases, it is not using the DINOv2 backbone [2026-07-07 05:38:48] [WARNING] rf-detr - Using a different number of positional encodings than DINOv2, which means we're not loading DINOv2 backbone weights. This is not a problem if finetuning a pretrained RF-DETR model. [2026-07-07 05:38:48] [WARNING] rf-detr - Using patch size 16 instead of 14, which means we're not loading DINOv2 backbone weights. This is not a problem if finetuning a pretrained RF-DETR model. 2. Can you simplify the training output? The terminal is very verbose and most of the time we need to use LLMs to understand the errors. Thank you

…

On Mon, Jul 6, 2026 at 3:03 PM Jesús Royeth ***@***.***> wrote: *JESUSROYETH* left a comment (roboflow/rf-detr#1186) <#1186 (comment)> Yes, for the YOLO format it works like you expected: you drop the background image into train/images (and valid/images) and you don't create any .txt for it in labels/. The loader then treats it as a background sample with 0 boxes. The reason is simple: the two formats have opposite defaults, and that is easy to miss. In COCO an image that is not registered in the json gets silently ignored — that is why in my first answer you had to add it to the images list. In YOLO it is the other way around: RF-DETR scans train/images directly and, for every image, it looks for a matching labels/<name>.txt. If that file is missing it just returns an empty target instead of skipping the image, an empty .txt (0 lines) works the same way. I checked this on rfdetr 1.8.3. I built a YOLO dataset with images in train/images, gave one a label file and left another with no .txt at all — len(dataset) counted both, and the unlabelled image returned boxes=(0, 4), labels=[] with no crash: background.jpg boxes=(0, 4) labels=[] # no .txt -> background empty_lbl.jpg boxes=(0, 4) labels=[] # empty .txt -> background obj.jpg boxes=(1, 4) labels=[0] One thing to keep in mind: the dataset still has to be a valid YOLO layout (a data.yaml/data.yml, plus train/ and valid/, each with images/ and labels/), otherwise RF-DETR won't detect it as YOLO in the first place. Versions: rfdetr 1.8.3, torch 2.12.1, torchvision 0.27.1, Python 3.12.3. Hope this helps! — Reply to this email directly, view it on GitHub <#1186?email_source=notifications&email_token=BSIM7DTEOEYJVNA77JD6CAT5DNW7TA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBZGEZDKMJTGU4KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4891251358>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BSIM7DSBJTI5RH6J7GSZYKD5DNW7TAVCNFSNUABFKJSXA33TNF2G64TZHM4TKMJVGM2DENRQHNEXG43VMU5TIOBQHE4DMNJYGYYKC5QC> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/BSIM7DQCJK2B2HOHDWKQK535DNW7TA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBZGEZDKMJTGU4KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android <https://github.com/notifications/mobile/android/BSIM7DQXIGPTKZMJVJENFFT5DNW7TA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBZGEZDKMJTGU4KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you were mentioned.Message ID: ***@***.***>