In computer vision applications, variations in lighting have long posed significant challenges for image annotation. This issue is particularly evident in high-precision scenarios such as autonomous driving and surveillance (involving humans and vehicles), as well as outdoor monitoring in agriculture. Complex shifts in lighting not only affect data quality itself, but also lead to biases in feature extraction and reduced confidence in annotations. Because lighting is a fundamental factor in image formation, any fluctuation in light intensity or color temperature directly influences the level of detail, contrast, and color accuracy in captured images. These effects manifest in several ways:
1. Fluctuations in Image Quality
Different lighting conditions can easily cause overexposed or underexposed images, resulting in lost detail and an overall blurred appearance. Overexposure may wash out object surface colors and textures, while underexposure renders the darker regions barely distinguishable—both circumstances undermine annotation accuracy. For instance, in nighttime or poorly lit environments, pedestrians or vehicles often become difficult to detect and accurately label.
Figure 1 Overexposed (left) and underexposed (right)
2. Impact of Shadows
Shadows can alter an object’s overall appearance or partially obscure it, making it difficult to identify outlines, shapes, and details. In real-world conditions, lane markings, pedestrians, vehicles or road signs may be partially covered by tree shade or building shadows, making it challenging for annotation systems to distinguish between genuine objects and shadow regions and thus increasing the risk of misclassification.
Figure 2 Pedestrians and vehicles covered by shadows
3. Color Distortion
The color temperature of light varies throughout the day and in different environments. For example, early morning or late afternoon light often takes on a reddish or yellowish hue, while cloudy or overcast conditions produce cooler tones. These shifts in color can interfere with detection and annotation algorithms that rely on color features, thereby reducing recognition accuracy.
Figure 3 Warm tones at dusk (left) and cool tones on a rainy day (right)
4. Changes in Contrast
Insufficient or overly strong light can result in abnormally low or high contrast levels, making it harder to distinguish the boundaries between objects and the background. Foggy or overcast conditions generally lower contrast, causing object outlines to become blurred; conversely, under intense direct light, excessively high contrast may cause some details to be “washed out” in bright areas.
Figure 4 Low contrast in fog (left) and excessively high contrast under strong light (right)
To illustrate how lighting variations pose potential challenges in actual production settings, this article highlights two representative scenarios — autonomous driving and agricultural monitoring — and examines how changes in lighting adversely affect the annotation of specific targets. Meanwhile, we also showcase AI-driven annotations from T-Rex Label, an AI image annotation tool, verifying that T-Rex Label can maintain stable and reliable annotation performance even under complex lighting conditions:
1. Autonomous Driving
Autonomous driving requires extremely high accuracy in image annotation, and even slight fluctuations in lighting can disrupt the algorithm’s perception of the environment, particularly in identifying pedestrians and vehicles:
a) Pedestrian Recognition
Under intense sunlight, pedestrians’ clothing or faces can be overexposed, making it difficult for detection algorithms to extract complete features and thereby affecting accurate identification of individuals. When pedestrians are in shadow, parts of their bodies may be obscured, complicating detection of their overall outline and increasing the risk of misidentification. In low-light or nighttime conditions, subtle pedestrian details are “swallowed” by darkness, further raising the difficulty of detection and potentially preventing the algorithm from accurately recognizing pedestrians in time.
b) Vehicle Recognition
Under strong light, reflective metal surfaces on vehicles may cause glare, concealing critical features and hindering algorithms’ ability to identify vehicles. In shadowed areas, partial vehicle obstruction results in incomplete outlines, complicating judgments about vehicle features. In nighttime or low-light environments, the contrast between vehicle headlights and surroundings might be low, causing the algorithm to confuse vehicles with the background and leading to detection errors.
Based on this, T-Rex Label’s AI annotation performance under various lighting conditions for pedestrians (in yellow bounding boxes) and vehicles (in red bounding boxes) is demonstrated as follows:
Figure 5 AI annotation results for overexposed (left) and underexposed (right) environments
Figure 6 AI annotation results in warm tones (left) and cool tones (right)
Figure 7 AI annotation results in high-contrast (left) and foggy low-contrast (right) environments
Figure 8 AI annotation results under shadow
2. Agricultural Monitoring
When robots or drones conduct aerial surveys over farmland or woodland, the captured images often face multiple lighting disturbances: Sunlight intensity changes with the weather and time of day, cloud movements cause certain regions to brighten or darken abruptly, and seasonal variations alter both the angle and color temperature of incident light. In such a dynamic and complex lighting environment, the color and texture of crops and trees can exhibit significant fluctuations. Consequently, as shown in Figure 9, certain areas may become overexposed or overly dark, making it difficult to identify and label relevant features accurately.
Figure 9 overexposure or shadow coverage
T-Rex Label’s AI annotation results for fruit recognition under various lighting changes are shown below:
Figure 10 Annotation results for overexposed fruit or fruit in shadow
Experimental findings confirm that although changes in lighting present substantial challenges for image annotation, T-Rex Label, bolstered by its powerful Zero-Shot Visual Prompt capabilities, can still maintain efficient and precise annotations across a diverse range of complex lighting conditions. As computer vision technology continues to mature and real-world production demands broaden, the challenges posed by lighting variations to image annotation tools will only grow more diverse. The race to master lighting adaptation is just beginning.
Appendix
T-Rex Label access (FREE!):https://www.trexlabel.com/?source=dds