Different results from detect vs val, "no detections"

[PeterKBailey]

Hello! I hope I'm asking this in the right place but I'm new to YOLO and have been trying to determine if I have properly built and structured my dataset.

I used the following command to first train the model on a very small data subset (222 images)

python yolov5/train.py --img 640 --batch 16 --epochs 3 --data dataset.yaml --weights '' --cfg yolov5n.yaml

I also included a validation directory which contains the exact same images in my training set. All I'm trying to do right now is see that I can make a model which overfits / perfectly predicts for this training data.
After having completed the training I did two things:

python detect.py --weights best.pt --img 640 --data dataset.yaml --source image.jpg --conf-thres 0.01
and
python .val.py --weights best.pt --img 640 --data dataset.yaml
with the following results: (output from detect on left, and output from val on the right)

Please excuse the resolution, this is a cropped in screenshot but is only meant to show that the validation batch worked as expected but that detection on the image apparently did not.

Things I have tried:

• I have tried lowering and lowering the conf-thres but nothing happens until suddenly the entire image is taken up by spurious bounding box predictions.
• I did a training run of 15 epochs, everything else the same, no difference in validation vs detection
• I did a training run using my dataset but beginning with the preset YOLOv5n weights, still no difference

I'm not sure where to go from here so advice would be greatly appreciated! Thanks!

OS: Windows 10
YOLOv5 v7.0-295-gac6c4383 Python-3.11.6 torch-2.2.1+cpu CPU

Edit with additional things I have tried:

• Replacing the input image to detect with a resized 640x640 image (both by squashing/stretching and by letterboxing)
• Training a yolov5s model instead (val still works, detect does not)
• setting conf-thres to 0.001 and iou to 0.6 when detecting (which are apparently the defaults for validation), no detections.
• Upped dataset size to 4130 training images with 950 validation images, still validation shows results but detection does not

I have also double checked my bounding boxes with the YoloBBoxChecker and am certain that my text files are correct. I didn't doubt this particularly given that val is successful but figured it was one more item on the list.