One shot object detector doesn't work for me

I'm having a problem with the one shot object detector. The model does not return a single prediction.

My setup:

I'm trying to detect two different kinds of colorimetric test strips (4mm width, 10cm height, negletible thickness). I have thousands of photos which show these test strips in a 2D setting (just a photo of them laying on a table from above). So I think the context is really 2D and the one shot detector should work for this scenario.

I tried to train the model in several ways, but I don't get any prediction back, only empty results.

I haven't found much documentation, I only watched the WWDC session. So in particular I don't know what kind of size restrictions apply to it or what a optimal object size in relation to the image size should be.

Any help appreciated.

Here is what I did:

train = tc.image_analysis.load_images('./train/mini')
train['label'] = train['path'].element_slice(13,-9)
>>> train                                                
Columns:
  path str
  image Image
  label str

Rows: 2

Data:
+---------------------------+-----------------------+-------+
|            path           |         image         | label |
+---------------------------+-----------------------+-------+
| ./train/mini/aaamini.jpeg | Height: 681 Width: 26 |  aaa  |
|  ./train/mini/bbmini.jpeg | Height: 707 Width: 27 |   bb  |
+---------------------------+-----------------------+-------+
[2 rows x 3 columns]

model = tc.one_shot_object_detector.create(train, 'label')
model                                                     
Class                                    : OneShotObjectDetector

Model summary
-------------
Number of classes                        : 2
Input image shape                        : (3, 416, 416)

Synthetic data summary
----------------------
Number of synthetically generated examples : 1902
Number of synthetically generated bounding boxes : 1902

Training summary
----------------
Training time                            : 2h 33m 49s
Training iterations                      : 7000
Training epochs                          : 117
Final loss (specific to model)           : 1.3897

test = tc.image_analysis.load_images('./test')

test                                          
Columns:
  path str
  image Image

Rows: 20

Data:
+-------------------------------+-------------------------+
|              path             |          image          |
+-------------------------------+-------------------------+
| ./test/1...                   | Height: 1008 Width: 756 |
...
+-------------------------------+-------------------------+
[20 rows x 2 columns]

result = model.predict(test)
Predicting  1/20
Predicting 20/20
>>> result
dtype: list
Rows: 20
[[], [], [], [], [], [], [], [], [], [], [], [], [], [], [], [], [], [], [], []]