Of all the past detections, birds were found at the red feeder 9% of the time, at the platform feeder 44% of the time, and at the suet feeder 46% of the time. "At" a feeder meant that the bounding box fell within the approximate areas shown below.
Each detetector's contribution to the overall number of detections is listed below. The detectors trained on automatically partitioned data also have descriptive images generated from the training data used for each detector by dlib. (These images were generated before I cleaned up the auto-partitioned data for each detector, so some images don't look like they belong. Also, the two "red birdfeeder" detectors were created mostly to detect birds hanging off the red feeder, and these were split from the same dlib-generated cluster.) I've named the detectors according to the rough orientation of birds they were trained on. One interesting thing I noticed when labeling images is that the birds face left far more frequently than they face right!
|Bird Faces: 8%||Sideways Facing Left: 1%||Diagonal Facing Left: 1%||Diagonal Facing Right: 5%|
|Up/Down Facing Left: 9%||Up/Down Facing Right: 3%||Bottom of Red Feeder, Facing Forward: 1%||Bottom of Red Feeder, Facing Sideways: 0%|
|Hand-partitioned, Updown: 43%||Hand-partitioned, Side: 8%||Hand-partitioned, Front/Face: 8%||Hand-partitioned, Back: 6%|
Otherwise this is a stub for now. More coming soon... I have a lot of analysis ideas but ran out of time.
Although it's not something I can calculate from the database, some readers may be interested to know what species of birds show up at my feeder. I looked through my past ebird records and this is what I've seen at or near the feeder:
And then there's sometimes this visitor. He showed up during one of my data collections and was duly added as a negative training example. :)