nehdf/Landmark_Navigation_Blind_Spot

Landmark-based drone navigation in African neighbourhoods

Submission to the Fatima Fellowship Blind Spots of Frontier Models challenge. All artifacts: hf://buckets/nehdf/Landmark_Navigation_Blind_Spot.

1. The blind spot, and why it matters to me

My evaluation investigates whether vision-language models can interpret spatial directions in aerial images when image orientation changes and navigation relies on local landmarks. Across 1,850 trials covering nine cities, cardinal-direction accuracy dropped from 78% in the standard condition to 2% after a 180° rotation, despite an explicit north indicator. Highlighting landmarks improved grid-position accuracy from 24% to 80%, while route-relative instructions achieved only 30%, close to the 25% chance level. These findings reveal potential weaknesses in orientation awareness, landmark recognition, and route-based spatial reasoning, which are important for drone navigation in areas without conventional street addresses.

Evaluation setup

Models: Qwen/Qwen3.5-4B @ 851bf6e806ef

Loading (Google Colab, Tesla T4, transformers 5.20.2): AutoProcessor + AutoModelForMultimodalLM (fallback AutoModelForImageTextToText), fp16/bf16, chat template with enable_thinking=False. Full code in notebook/.

Scoring. No free-text parsing. After the prompt, the text Answer: is appended and every option (grid cell or marker letter) is scored by its summed token log-probability as a continuation (teacher forcing). The answer is the most probable option; probabilities renormalised over the options give p_gold and a confidence. A separate format probe (60 free generations, 384 tokens) measures whether the model follows the required answer format at all. A first run that parsed free text (v3) was discarded: the model ignored the format and a lenient parser read the article "a" as option A.

Data. Drone imagery from the Open Cities AI Challenge (GFDRR, CC BY 4.0), tiles of 250 m rendered at 768 px (~0.33 m/px); building footprints from the same dataset; landmarks and roads from OpenStreetMap (ODbL). Only residential tiles are kept (>= 30 buildings, median footprint <= 350 m2, >= 99% valid pixels).

city	tiles	cases
Accra, Ghana	25	51
Dar es Salaam, Tanzania	27	47
Kampala, Uganda	6	13
Kinshasa, DR Congo	2	3
Monrovia, Liberia	2	4
Ngaoundéré, Cameroon	15	22
Niamey, Niger	11	17
Pointe-Noire, Congo	3	4
Zanzibar, Tanzania	2	2

Planned-ness. Each tile gets a layout_score = mean of the standardised building orientation order (entropy of footprint orientations mod 90°) and road orientation order φ (Boeing 2019). Tertiles define planned / mixed / unplanned.

Tasks (gold answers computed geometrically): L1 grid localisation of a landmark; L2 cardinal relation; L3 route midpoint; L4 route-relative delivery (behind / in front / left / right when coming from another landmark). 163 cases, 1630 trials per model over conditions blind, standard, oracle, rot90, rot180 and languages en, fr. blind = no image; oracle = referenced landmark outlined; rot90/rot180 = image and compass rotated.

Results

model_id	condition	layout	n	accuracy	ci	p_gold	chance
Qwen/Qwen3.5-4B	blind	mixed	64	0.19	[0.11, 0.30]	0.19	0.17
Qwen/Qwen3.5-4B	blind	planned	166	0.14	[0.10, 0.21]	0.16	0.18
Qwen/Qwen3.5-4B	blind	unplanned	96	0.18	[0.11, 0.27]	0.18	0.19
Qwen/Qwen3.5-4B	oracle	mixed	64	0.81	[0.70, 0.89]	0.56	0.17
Qwen/Qwen3.5-4B	oracle	planned	166	0.67	[0.60, 0.74]	0.51	0.18
Qwen/Qwen3.5-4B	oracle	unplanned	96	0.76	[0.67, 0.83]	0.56	0.19
Qwen/Qwen3.5-4B	rot180	mixed	64	0.03	[0.01, 0.11]	0.11	0.17
Qwen/Qwen3.5-4B	rot180	planned	166	0.16	[0.11, 0.23]	0.19	0.18
Qwen/Qwen3.5-4B	rot180	unplanned	96	0.09	[0.05, 0.17]	0.15	0.19
Qwen/Qwen3.5-4B	rot90	mixed	64	0.19	[0.11, 0.30]	0.16	0.17
Qwen/Qwen3.5-4B	rot90	planned	166	0.23	[0.18, 0.30]	0.20	0.18
Qwen/Qwen3.5-4B	rot90	unplanned	96	0.18	[0.11, 0.27]	0.19	0.19
Qwen/Qwen3.5-4B	standard	mixed	64	0.53	[0.41, 0.65]	0.27	0.17
Qwen/Qwen3.5-4B	standard	planned	166	0.43	[0.36, 0.51]	0.27	0.18
Qwen/Qwen3.5-4B	standard	unplanned	96	0.48	[0.38, 0.58]	0.30	0.19

p_gold = mean probability the model puts on the correct option.

By instruction style and condition:

model_id	instruction_style	condition	n	accuracy	ci
Qwen/Qwen3.5-4B	absolute	oracle	114	0.82	[0.73, 0.88]
Qwen/Qwen3.5-4B	absolute	rot180	114	0.17	[0.11, 0.25]
Qwen/Qwen3.5-4B	absolute	rot90	114	0.16	[0.10, 0.24]
Qwen/Qwen3.5-4B	absolute	standard	114	0.22	[0.15, 0.30]
Qwen/Qwen3.5-4B	cardinal	oracle	138	0.91	[0.85, 0.95]
Qwen/Qwen3.5-4B	cardinal	rot180	138	0.00	[0.00, 0.03]
Qwen/Qwen3.5-4B	cardinal	rot90	138	0.12	[0.07, 0.18]
Qwen/Qwen3.5-4B	cardinal	standard	138	0.78	[0.71, 0.84]
Qwen/Qwen3.5-4B	relative	oracle	40	0.15	[0.07, 0.29]
Qwen/Qwen3.5-4B	relative	rot180	40	0.12	[0.05, 0.26]
Qwen/Qwen3.5-4B	relative	rot90	40	0.45	[0.31, 0.60]
Qwen/Qwen3.5-4B	relative	standard	40	0.17	[0.09, 0.32]
Qwen/Qwen3.5-4B	route	oracle	34	0.35	[0.21, 0.52]
Qwen/Qwen3.5-4B	route	rot180	34	0.41	[0.26, 0.58]
Qwen/Qwen3.5-4B	route	rot90	34	0.47	[0.31, 0.63]
Qwen/Qwen3.5-4B	route	standard	34	0.35	[0.21, 0.52]

Paired comparisons:

model	comparison	pairs	diff	p (exact McNemar)
Qwen/Qwen3.5-4B	image_vs_blind	326	+0.30	5.05e-20
Qwen/Qwen3.5-4B	oracle_vs_standard	326	+0.26	1.35e-16
Qwen/Qwen3.5-4B	rot90_vs_standard	326	-0.26	9.93e-13
Qwen/Qwen3.5-4B	rot180_vs_standard	326	-0.35	1.4e-21
Qwen/Qwen3.5-4B	fr_vs_en	163	-0.09	0.00434

Results summary (auto-generated from the data)

Auto-generated from results/ on 2026-10-21; numbers only, no interpretation. Answers are scored by option likelihood (no free-text parsing).

• Scale: 163 cases from 93 residential drone-image tiles in 9 cities (Accra, Ghana; Dar es Salaam, Tanzania; Kampala, Uganda; Kinshasa, DR Congo; Monrovia, Liberia; Ngaoundéré, Cameroon; Niamey, Niger; Pointe-Noire, Congo; Zanzibar, Tanzania); 1630 scored trials.

Qwen/Qwen3.5-4B

• Standard condition: accuracy 46.6% (chance 18.4%); mean probability on the correct option 27.6%.
• Without the image: 16.3% vs 46.6% with it (significant, p = 5.05e-20).
• Planned vs unplanned (accuracy): 43.4% vs 47.9%, gap -4.5 points, 95% CI [-29.4, +4.5] (includes zero).
• Planned vs unplanned (probability on the correct option): 26.8% vs 29.8%, gap -3.0 points, 95% CI [-14.7, +0.9] (includes zero).
• Logistic regression with city fixed effects: each SD of planned-ness multiplies the odds of a correct answer by 1.08 [0.74, 1.58] (not significant, p = 0.68).
• Linear regression on probability of the correct option with city fixed effects: -0.5 points per SD of planned-ness [-2.6, +1.6] (not significant, p = 0.62).
• Outlining the landmark (oracle): 46.6% -> 72.7% (significant, p = 1.35e-16).
• Absolute instructions (n=114 per condition): standard 21.9%, oracle 81.6%, rot90 15.8%, rot180 16.7%.
• Cardinal instructions (n=138 per condition): standard 78.3%, oracle 91.3%, rot90 11.6%, rot180 0.0%.
• Route instructions (n=34 per condition): standard 35.3%, oracle 35.3%, rot90 47.1%, rot180 41.2%.
• Relative instructions (n=40 per condition): standard 17.5%, oracle 15.0%, rot90 45.0%, rot180 12.5%.
• Rotating the image (rot90): 46.6% -> 20.9% (significant, p = 9.93e-13).
• Rotating the image (rot180): 46.6% -> 11.7% (significant, p = 1.4e-21).
• Answer unchanged after rotation (mapped back to the original frame): rot180 18.4%, rot90 28.5%.
• English vs French: 50.9% vs 42.3% (significant, p = 0.00434).
• Lowest city (n >= 20): Accra, Ghana 43.1% (n=102); highest: Niamey, Niger 64.7% (n=34).
• Grid tasks: exact 21.9%; within one cell 51.8%.
• Calibration: mean confidence 37.3% vs accuracy 46.6%; expected calibration error 0.12; confidently wrong (confidence >= 70%) on 2.1% of trials.
• Format probe (free generation, n=60): followed the required 'Answer: X' format in 53.3% of trials; hit the 384-token limit in 46.7%; median length 378 tokens; free-text accuracy 18.3%.

3. A path forward

I would develop an orientation-aware training framework using open aerial imagery, verified geographic landmarks, and locally collected navigation instructions. Random image rotations with correctly updated compass indicators would help test and reduce the assumption that north is always at the top. A curriculum combining landmark localization, cardinal directions, and route-relative reasoning could be implemented through parameter-efficient fine-tuning. I would evaluate improvements in rotation consistency, landmark localization, and navigation accuracy, using geographically separate test cities to assess generalization. As an immediate baseline, I would also compare predictions using the original images against north-aligned images and explicit numerical drone headings.

Limitations (factual)

• Gold answers come from OpenStreetMap geometry; the imagery dates from 2018–2020 and OSM is more recent, so some landmarks may be absent or moved. Cases can be screened through case_overrides.csv.
• Likelihood scoring measures which option the model prefers when forced to answer; the format probe reports separately how often it would have produced a usable answer on its own.
• Cities contribute unequal numbers of cases and layouts; regressions include city fixed effects when estimable.
• Landmark categories are referred to by type only; some (schools, churches) are hard to recognise from above, which the oracle condition helps separate from reasoning errors.
• layout_score is a proxy for planned-ness, built from footprint and road orientation.

Files

Path	Content
data/annotations.csv	final cases (auto-generated + author overrides)
data/case_overrides.csv	author QC decisions and local phrasings
data/images/, data/rendered/	tiles (PNG), and the exact image shown in each trial (JPEG)
data/tiles.json	tile metadata, landmarks, planned-ness scores
results/option_scores.jsonl	log-probability of every option for every trial, with model commit
results/format_probe.jsonl	free generations of the format probe
results/ (other)	scored trials, aggregated tables, calibration curve, summary.json, failure review
figures/	figures and failure cards
notebook/	the notebook that produced everything
model/model_reference.json	model ids, commits, decoding settings

Sources and licences

• Imagery & footprints: GFDRR Labs (2020). Open Cities AI Challenge Dataset, v1.0. doi:10.34911/rdnt.f94cxb — imagery CC BY 4.0.
• Landmarks & roads: © OpenStreetMap contributors, ODbL 1.0.
• Derived annotations in this bucket are released under ODbL 1.0 (share-alike from OSM); images remain CC BY 4.0.

Reproduce

from huggingface_hub import sync_bucket
sync_bucket("hf://buckets/nehdf/Landmark_Navigation_Blind_Spot", "./landmark_nav")

Open notebook/ in Colab, set PROJECT_DIR, run all cells (cached sources are reused).