2025-03-07 Weedbot Report 1
Statistics based on pixels
Asya rezultāti
| Recall | Precision | pr2 | |
|---|---|---|---|
| carrots | 0.933 | 0.444 | 0.935 |
| weeds | 0.788 | 0.266 |
Weedbot rezultāti (nepieciešams izpētīt atšķirību pēc atsūtītā pirmkoda)
Calculation for pr2:
11pr2 = np.sum(TP_carrot) / np.sum(TP_carrot + FP_carrot_where_actual_weed)Problem with ignoring the false-positive carrot predictions where its actually ground is that its ambiguous - any prediction outside would count, and since the blobs of weeds are very small even bad carrot predictions would produce a good score.
Statistics based on instances
Explanation (actual result table below)
In the image below (true to the left, predicted on the right).
Total blobs: 9 in true mask, 8 in predicted
Total 9 marked weed blobs, 7 of those predicted correctly.
Total 8 weed blobs predicted, with 1 FP.
Results table (Instance based)
| Metric | Value |
|---|---|
| Total Images | 92 |
| Total Marked Weed Blobs | 397 |
| Total Predicted Weeds | 529 |
| Correct Predictions | 367 |
| Missed Weeds (FN) | 30 |
| False Positive Weeds (FP) | 162 |
| Recall | 92.44% |
| Precision | 69.38% |
| False Positive Weeds on Carrots | 5.4% (5/92) |
Most dangerous cases - there is carrot, which is not detected, and on top of it a weed is detected, even though it shouldn't be there.
11metric = (true_carrot == 1) * (pred_carrot == 0) * (true_weeds == 1) * (pred_weeds == 0)Detected 5.4% (5/92) images where a carrot would be incorrectly lasered with individual blob sizes larger than 56 pix. Each image had one such spot.
Examples
Top left - original image resized to prediction 550x550
Top right - the dangerous area where a carrot would be lasered
Bottom left - true masks of carrots and weeds.
Bottom right - predicted masks of carrots and weeds.
