UrbanInsects

[qgeissmann]

https://figshare.com/articles/figure/An_annotated_image_dataset_of_urban_insects_for_computer_vision_and_deep_learning/28280792
Since we can use bboxes in training (#158).

[qgeissmann]

UrbanInsects — recovering unannotated insects with flat-bug

1,148 insects were present in the images but missing from the published annotation. They have been reviewed one by one and added to CVAT task 2567343, taking it from 24,758 to 25,906 boxes — a 4.6% increase.

How they were found

flat-bug was run over all 69 cards. Every detection that did not overlap an existing annotation (<15% of its area inside any annotated box), was at least 32 px across, and did not touch the card border became a candidate. Each candidate was then judged by hand in a review UI, so nothing here is model output taken on trust.

	count	share
flat-bug detections not matching an annotation	3,273	100%
→ confirmed insects (rescued)	1,148	35%
→ false positives (rejected)	2,125	65%

So roughly one in three of flat-bug's unannotated detections was a real insect. The other two thirds were debris, glue texture, card printing and fragments of already-annotated specimens.

What was missed

Forty rescued detections, spanning the confidence range from 0.35 to 0.96. All are unambiguous insects that the original annotation omitted.

The missed insects are smaller

longest side (px)	originally annotated	rescued	rejected
p10	66	55	52
p25	91	63	61
p50	136	80	89
p75	194	103	138
p90	251	145	214
mean	159	93	120

The median rescued insect is 80 px across against 136 px for those already annotated, and 64% of rescued instances fall below the first quartile of the annotated size distribution. The original annotators were evidently working down to a size and stopping.

Confidence is a much weaker signal: mean 0.76 for rescued against 0.64 for rejected. Thresholding on confidence alone would not have separated them, which is why every candidate was reviewed.

Per site

site	cards	original	reviewed	rescued	increase
LUZI	18	13,494	1,294	385	+2.9%
PONTIAN	24	6,187	828	242	+3.9%
TEBRAU	15	1,696	418	202	+11.9%
SENAI	6	2,285	445	164	+7.2%
ISKANDARPUTERI	6	1,096	288	155	+14.1%
total	69	24,758	3,273	1,148	+4.6%

Per card

All 69 cards, ordered by how many insects were recovered.

card	original	reviewed	rescued	rejected	increase
ISKANDARPUTERI_170824FT2_C100	199	92	64	28	+32.2%
LUZI_150824_A315	1,490	138	53	85	+3.6%
LUZI_150824_B316	2,035	126	50	76	+2.5%
PONTIAN_170824_A119	333	77	50	27	+15.0%
ISKANDARPUTERI_170824FT2_B099	279	59	40	19	+14.3%
ISKANDARPUTERI_170824FT2_A098	201	69	39	30	+19.4%
LUZI_150824_A300	182	75	38	37	+20.9%
SENAI_050824_B010	531	101	38	63	+7.2%
SENAI_050824_C008	245	62	35	27	+14.3%
SENAI_050824_C011	339	127	31	96	+9.1%
TEBRAU_160824_B171	166	53	31	22	+18.7%
LUZI_150824_B310	82	109	28	81	+34.1%
LUZI_150824_C311	99	80	27	53	+27.3%
SENAI_050824_A006	265	41	27	14	+10.2%
LUZI_150824_B313	720	86	26	60	+3.6%
LUZI_150824_B301	294	62	22	40	+7.5%
PONTIAN_170824_A107	234	36	22	14	+9.4%
LUZI_150824_A309	102	98	20	78	+19.6%
PONTIAN_170824_C121	512	65	20	45	+3.9%
SENAI_050824_A009	521	92	20	72	+3.8%
LUZI_150824_C317	1,920	57	19	38	+1.0%
LUZI_150824_C314	454	93	18	75	+4.0%
TEBRAU_160824_A164	79	23	18	5	+22.8%
TEBRAU_160824_A170	198	34	18	16	+9.1%
PONTIAN_170824_C109	229	22	17	5	+7.4%
LUZI_150824_C302	205	55	16	39	+7.8%
TEBRAU_160824_B165	99	23	16	7	+16.2%
TEBRAU_160824_C172	116	25	16	9	+13.8%
LUZI_150824_A303	83	34	15	19	+18.1%
TEBRAU_160824_A173	201	27	15	12	+7.5%
PONTIAN_170824_A101	68	55	14	41	+20.6%
PONTIAN_170824_A110	185	28	14	14	+7.6%
LUZI_150824_C305	107	43	13	30	+12.1%
PONTIAN_170824_C112	354	28	13	15	+3.7%
SENAI_050824_B007	384	22	13	9	+3.4%
TEBRAU_160824_B168	88	21	13	8	+14.8%
LUZI_150824_A312	588	121	12	109	+2.0%
TEBRAU_160824_A158	87	31	12	19	+13.8%
TEBRAU_160824_B174	206	31	12	19	+5.8%
TEBRAU_160824_C175	110	30	12	18	+10.9%
PONTIAN_170824_A104	71	32	11	21	+15.5%
LUZI_150824_B307	1,915	56	10	46	+0.5%
PONTIAN_170824_A122	403	70	10	60	+2.5%
TEBRAU_160824_C166	65	18	10	8	+15.4%
LUZI_150824_B304	109	36	9	27	+8.3%
PONTIAN_170824_B105	67	44	9	35	+13.4%
TEBRAU_160824_C169	83	18	9	9	+10.8%
PONTIAN_170824_B120	579	72	8	64	+1.4%
TEBRAU_160824_A167	78	13	8	5	+10.3%
TEBRAU_160824_C160	58	24	8	16	+13.8%
PONTIAN_170824_B111	323	27	7	20	+2.2%
PONTIAN_170824_C106	68	40	7	33	+10.3%
ISKANDARPUTERI_170824FT1_A095	134	27	6	21	+4.5%
LUZI_150824_C308	1,631	10	5	5	+0.3%
PONTIAN_170824_A113	419	20	5	15	+1.2%
PONTIAN_170824_B108	322	16	5	11	+1.6%
PONTIAN_170824_B123	595	42	5	37	+0.8%
PONTIAN_170824_C103	84	39	5	34	+6.0%
ISKANDARPUTERI_170824FT1_C097	130	22	4	18	+3.1%
LUZI_150824_A306	1,478	15	4	11	+0.3%
PONTIAN_170824_B102	95	54	4	50	+4.2%
PONTIAN_170824_B114	398	19	4	15	+1.0%
PONTIAN_170824_C124	382	11	4	7	+1.0%
TEBRAU_160824_B159	62	47	4	43	+6.5%
PONTIAN_170824_C118	37	7	3	4	+8.1%
ISKANDARPUTERI_170824FT1_B096	153	19	2	17	+1.3%
PONTIAN_170824_A116	82	7	2	5	+2.4%
PONTIAN_170824_B117	71	6	2	4	+2.8%
PONTIAN_170824_C115	276	11	1	10	+0.4%

Per-card increase ranges from +0.3% to +34.1% (median +7.5%), so the under-annotation is spread across the corpus rather than confined to a few bad cards.

Caveats

• Candidates were thresholded at confidence ≥0.35 and ≥32 px, so insects fainter or smaller than that were never offered for review. 1,148 is a lower bound.

• Detections overlapping an existing annotation by ≥15% were filtered out, which also removes genuine cases where two touching specimens were annotated as one.

• Rescued boxes are flat-bug's, not hand-drawn: they were confirmed to contain an insect, but their exact extent was not adjusted.

---

Data: figshare 10.6084/m9.figshare.28280792.v1 (CC BY 4.0). Annotations: CVAT task 2567343 UrbanInsects-bbox.