基于YOLOv11与Roboflow的MoodAI Player面部情绪检测项目训练图像采集优化咨询

Hi everyone,

I'm working on a university class project called MoodAI Player—a small computer vision project that detects facial emotions via a camera and triggers music/LED responses on a Raspberry Pi. I'm planning to use YOLOv11 for model training, with Roboflow handling data annotation. My dataset covers 6 emotion classes: happy, sad, angry, neutral, fearful, sleepy, with roughly 100 images per class (totaling around 600 images).

Right now, I suspect my current image collection strategy might be the bottleneck for the project, so I'm reaching out for targeted advice on how to refine this part. Here's what my current collection plan looks like:

• Prioritize front-facing faces
• Use clean, uncluttered backgrounds
• Capture images in good lighting conditions
• Keep faces occupying 60%-80% of the frame
• Include moderate variation: slight shooting angle changes, different lighting environments, and faces of different people
• Split each class into 50 images of people wearing glasses, and 50 without

I have four specific questions I'd love to get answers to:

1. Should I stick with clean, polished images, or should I incorporate more real-world "noise" like cluttered backgrounds, varying shooting distances, and less ideal framing?
2. How much scene diversity is actually practical and useful for my use case? Do I need to include side-profile angles, low-light/shadow scenarios, or partial occlusions (like hair covering parts of the face, hands near the face)?
3. Even with Roboflow's built-in data augmentation features, is 100 images per class still too sparse to get reasonable real-time performance? (Note: This is a class project, so I'm not aiming for state-of-the-art accuracy—just a model that works decently in real scenarios.)
4. What common mistakes should I avoid when building an emotion-focused dataset, especially when dealing with easily confused classes like sleepy vs. neutral?

Any tips, insights, or personal experiences related to these questions would be a huge help. Thanks in advance!

内容的提问来源于stack exchange，提问作者Aneeq Shakeel