NTU Outdoor Dataset

The existing public Person Re-ID datasets have a very limited number of cameras ranging from 6 to 15, reducing variation and diversity. As a result, recently proposed algorithms can achieve over 90% rank-1 accuracy on Market1501 and DukeMTMC-reID. However, real-world surveillance systems usually consist of over hundreds of cameras. Creating a more realistic dataset is the foundation for developing a more generalized and robust Person ReID model.

↓ Download NTU-Outdoor-38

Dataset Collection

The NTU-Outdoor dataset was collected within the NTU campus using actual surveillance cameras installed on lamp posts. There are a total of 34 camera groups, each containing two to four cameras pointing in different directions.

Figure 1: Camera Group Locations across the NTU campus.

Based on all 34 camera groups, 8 paths were designed for participants to walk past the cameras. The 8 paths contain 51 cameras from 23 camera groups. A total of 332 NTU students, staff and residents participated in this dataset collection.

Figure 2: Eight Paths for Data Collection — Figure 2: Eight designed paths for data collection across the NTU campus.

To increase annotation efficiency, a mobile web app was developed for this collection. By running the app on a smartphone, the GPS information and timestamps of the participant passing each camera were automatically recorded. This significantly reduced the searching time window for annotators.

Figure 3: NTU Outdoor Data Collection Web App Design.

After phase 1 annotation, 26,175 three-minute video clips were extracted. From those clips, a total of 45,397 bounding box images were annotated with 40 additional attribute labels, collected from 278 different people (unique identities) with 805 different appearances over an 8-week period.

Why This Dataset Matters

1. More Cameras

All existing large-scale datasets use only 6–15 cameras. The NTU-Outdoor dataset uses 51 real-world surveillance cameras covering the entire 2 km² NTU campus.

Dataset	Market-1501	DukeMTMC-reID	MSMT17	NTU Outdoor
Cameras	6	8	15	51

2. Actual Surveillance Viewing Angles

Market1501 and MSMT17 use cameras mounted on tripods, giving a near-horizontal view. Real surveillance cameras are mounted on lamp posts or ceilings with wide-angle, top-down views. NTU-Outdoor uses only actual lamp post surveillance cameras.

Figure 4: Viewing Angles comparison — Figure 4: Viewing Angles of Market1501, MSMT15 and NTU-Outdoor Dataset.

3. Day & Night Coverage

Real-world surveillance runs 24/7. All existing public datasets only use daytime video. NTU-Outdoor records all cameras for 24 hours non-stop, with 38.6% of images captured during nighttime.

Figure 5: Same person in afternoon, evening and night — Figure 5: Same Person in the Afternoon, Evening and Night Time in NTU-Outdoor Dataset.

Figure 6: Day and Night time distribution — Figure 6: Percentage of Day Time vs Night Time Images in NTU-Outdoor Dataset.

4. Attribute Labels

Appearance attributes are extremely valuable auxiliary information for training generalized Person Re-ID models. In NTU-Outdoor Dataset, most attributes are submitted by participants themselves, eliminating manual annotation and providing more accurate labels.

Figure 7: Upper and Lower Body Colour Distribution of NTU-Outdoor Dataset.

Attributes include: 5 upper body clothing types (t-shirt, polo-shirt, shirt, jacket, dresses), 5 lower body clothing types (shorts, jeans, pants, shirt, dresses), plus accessories (hat, glasses, handbag, backpack, messenger bag), and transportation (bike, e-scooter).

Figure 8: Distribution of Clothing Types and Accessories — Figure 8: Distribution of Clothing Types, Accessory and Transport in NTU-Outdoor Dataset.

Figure 9: Male and Female Ratio of Participants and Images in NTU-Outdoor Dataset.

Dataset	Market-1501	DukeMTMC-reID	MSMT17	NTU Outdoor
Attribute Labels	30	23	—	40

Overall Comparison

Dataset	Market-1501	DukeMTMC-reID	MSMT17	NTU-Outdoor	NTU-Outdoor-38 (Released)
Surveillance Camera	No	Yes	No	Yes	Yes
Number of Cameras	6	8	15	51	38
Collection Period	1 Day	1 Day	4 Days	8 Weeks	8 Weeks
Time Coverage	—	—	Morning / Noon / Afternoon	24 Hours (Day & Night)	24 Hours (Day & Night)
Number of Identities	1501	1812	4101	805	549
Number of BBoxes	32,668	36,411	126,441	66,084	48,347
Attribute Labels	30	23	—	40	40
Person Detection	DPM	DPM	Faster RCNN	YOLO V3	YOLO V3

↓ Download NTU-Outdoor-38 Dataset