kaihon/sg-aerial-scene-analyser-lora

SG Aerial Scene Analyser (LoRA)

QLoRA adapter for Qwen2.5-VL-7B-Instruct, fine-tuned for structured analysis of nadir (top-down) aerial imagery of Singapore.

Demo: kaihon/sg-aerial-scene-analyser

Usage

import torch
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from peft import PeftModel
from qwen_vl_utils import process_vision_info

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2.5-VL-7B-Instruct", torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(model, "kaihon/sg-aerial-scene-analyser-lora")
model.eval()

processor = AutoProcessor.from_pretrained(
    "Qwen/Qwen2.5-VL-7B-Instruct", min_pixels=256*28*28, max_pixels=1280*28*28
)

messages = [
    {"role": "system", "content": "You are an aerial scene analyst specialising in Singapore urban landscapes. Given a nadir (top-down) aerial image, return a JSON object with exactly these fields:\n{\n  \"caption\": \"3-5 sentences describing what is VISIBLE in a neutral surveyor tone, using Singapore-specific vocabulary (HDB block, hawker centre, covered walkway, MRT station). Name types not instances (MRT station not Bishan MRT). Only name globally unique landmarks (Marina Bay Sands, Jewel Changi Airport).\",\n  \"scene_type\": \"residential_hdb | commercial | industrial | port_terminal | airport | park_green | construction | mixed_use | transport\",\n  \"objects\": [{\"type\": \"hdb_block | condo | landed_house | shophouse | hawker_centre | mrt_station | bus_interchange | shopping_mall | warehouse | container_crane | cargo_ship | aircraft | construction_crane | sports_facility | place_of_worship | school\", \"count\": N}],\n  \"infrastructure\": [\"expressway | mrt_track | bus_lane | pedestrian_bridge | covered_walkway | park_connector | jetty | runway | taxiway\"],\n  \"terrain\": [\"water | urban | industrial | parkland | reclaimed_land | forest_reserve\"]\n}\nReturn ONLY the JSON object, no markdown fences or commentary."},
    {"role": "user", "content": [
        {"type": "image", "image": "path/to/aerial_image.jpg"},
        {"type": "text", "text": "Analyse this nadir aerial image of Singapore."},
    ]},
]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
img_inputs, _ = process_vision_info(messages)
inputs = processor(text=[text], images=img_inputs, padding=True, return_tensors="pt").to(model.device)

with torch.no_grad():
    gen = model.generate(**inputs, max_new_tokens=512, do_sample=False)
gen_trimmed = gen[:, inputs["input_ids"].shape[1]:]
output = processor.batch_decode(gen_trimmed, skip_special_tokens=True)[0]
print(output)

Output Format

The model returns a JSON object with 5 fields:

Field	Type	Description
caption	string	3-5 sentence description using Singapore-specific vocabulary
scene_type	string	One of 9 categories: residential_hdb, commercial, industrial, port_terminal, airport, park_green, construction, mixed_use, transport
objects	array	Detected objects with type and count
infrastructure	array	Infrastructure elements (e.g. expressway, mrt_track, covered_walkway)
terrain	array	Terrain types (e.g. urban, water, parkland)

Training

Parameter	Value
Base model	Qwen2.5-VL-7B-Instruct
Method	QLoRA (rank 8, alpha 16, dropout 0.05)
Training images	90 nadir aerial images of Singapore
Trainer	SFTTrainer (TRL)
Augmentation	On-the-fly random rotation + horizontal flip
Learning rate	1e-4
Batch size	2 (gradient accumulation 4, effective batch 8)
Precision	bf16
Early stopping	Patience 3 (on validation loss)
Final train loss	0.721
Hardware	NVIDIA L4 (23.7 GB), ~94 minutes

Evaluation (17-sample held-out test set)

Metric	Baseline (Qwen2.5-VL-7B)	Fine-tuned	Delta
Schema Compliance	100%	100%	—
Scene Type Accuracy	52.9%	70.6%	+17.7%
ROUGE-1 F1	0.325	0.536	+0.211
ROUGE-2 F1	0.040	0.268	+0.228
ROUGE-L F1	0.193	0.402	+0.209
BERTScore F1	0.875	0.917	+0.042
Object Mention F1	0.309	0.471	+0.161

Fine-tuning improves scene classification by +17.7%, doubles caption quality (ROUGE-L +0.21), and boosts object detection (F1 +0.16). Results are from a 70/15/15 train/val/test split; the deployed adapter is retrained on the full dataset (90 train / 16 val) for maximum coverage.

Framework Versions

• PEFT 0.18.1
• Transformers >= 4.49
• TRL (SFTTrainer)
• PyTorch (bf16)