如何提升货架商品品牌检测任务中Faster R-CNN模型的召回率？

Hey Caroline, great job getting your Faster R-CNN model up and running for shelf brand detection—let’s dive into actionable steps to boost that recall rate, especially at your 95% confidence threshold. Here’s what I’d recommend based on your setup:

Dataset & Annotation Improvements #

• Expand dataset diversity & augment strategically
  Your 300-image dataset is on the smaller side for 15 classes, so adding more variation will help the model generalize better. Try targeted data augmentation for shelf scenarios:
  • Simulate real-world conditions: random brightness/contrast shifts, perspective warps (to mimic different viewing angles), partial occlusion (cover parts of logos to replicate crowded shelves), and horizontal flips (avoid vertical flips unless your data includes upside-down brand cases).
  • Prioritize underperforming classes: If certain brands have lower recall, collect or generate (via synthetic overlays) more images of those logos in diverse shelf contexts.
• Audit & refine annotations
  Low recall often stems from missed or inaccurate annotations. Go through your validation set to check which true positive brands the model fails to detect:
  • Add annotations for any overlooked logos in training data.
  • Adjust tight/misaligned boxes to fully encompass brand logos—small annotation errors can break the model’s ability to learn consistent features.
• Fix class balance gaps
  Even with 100+ boxes per class, some brands might only appear in a tiny subset of images. Use weighted sampling during training to give more weight to underrepresented classes, forcing the model to prioritize learning their features.

Training Parameter Tuning #

• Adjust learning rate scheduling
  Your current conservative learning rate (1e-5 → 5e-6 at 200k steps) might be limiting the model’s learning potential. Try these tweaks:
  • Test a slightly higher initial rate (e.g., 1e-4) with a warmup phase: Ramp up from a tiny value to 1e-4 over the first 5k-10k steps to stabilize initialization without catastrophic forgetting.
  • Use dynamic scheduling: Replace the single drop with cosine annealing or step decay (e.g., halve the rate every 100k steps) to keep the model learning longer without plateauing.
• Extend training or add fine-tuning
  You stopped at 400k steps when total loss dropped below 0.1, but check your validation metrics: If validation recall is still improving, keep training with 5e-6 for another 100k-200k steps. If validation loss rises, switch to a tiny rate (5e-7) for fine-tuning to capture subtle features without overfitting.
• Diagnose with a lower confidence threshold
  Your 95% threshold is strict—temporarily drop it to 80% or 85%. If recall jumps significantly, the model is failing to assign high confidence to true positives, so we need to focus on making it more confident in correct detections rather than just filtering.

Model Structure & Configuration Tweaks #

• Switch to a multi-scale backbone
  The default Inception backbone might struggle with varying logo sizes on shelves. Try a backbone with Feature Pyramid Network (FPN) support (e.g., ResNet50-FPN or ResNet101-FPN). FPN fuses high-level semantic features with low-level spatial details, making it far better at detecting small/medium logos that single-scale backbones miss.
• Customize RPN anchors
  COCO’s default anchors aren’t optimized for shelf logos. Analyze your dataset’s annotation boxes: calculate average width/height and common aspect ratios (e.g., 1:1, 2:1, 1:2 for logos). Update your RPN anchor configuration to match these values—this helps the RPN generate relevant candidate boxes, reducing true positive misses.
• Optimize positive/negative sample balance
  The default 1:3 positive-to-negative ratio can prioritize background over rare/small logos. Use Online Hard Example Mining (OHEM) to let the model focus on hard-to-detect samples (e.g., partially hidden logos) during training, helping it learn to recognize challenging instances that drag down recall.

Post-Processing Adjustments #

• Tweak NMS settings
  A strict Non-Maximum Suppression (NMS) threshold (e.g., 0.5) might remove valid overlapping brand detections. Try increasing it to 0.6 or 0.7—this keeps more candidate boxes, boosting recall without a huge precision hit (you can refine this later).
• Use multi-scale inference
  During testing, run inference on multiple scaled versions of each image (e.g., 0.8x, 1.0x, 1.2x) and merge results. This helps the model catch logos that are too small/large in the original scale, improving recall for size-variable targets.

内容的提问来源于stack exchange，提问作者Caroline Wang