detex

Explaining object detectors

Installation

GPU

conda env create -n detex -f conda_gpu.yml
conda activate detex

CPU

conda env create -n detex_cpu -f conda_cpu.yml
conda activate detex_cpu

Download data and annotations

chmod +x scripts/download_cocoval.sh
./scripts/download_cocoval.sh

Running

Kernel SHAP

Example command: Run Kernel SHAP for the first 100 COCO images, using 2000 samples for approximating SHAP and assign for each perturbed feature a value of 0.5 (light gray), and store results.

python scripts/kernelshap_ssd.py \
    --first-images=100 \
    --batch-size=16 \
    --shap-samples=2000 \
    --baseline-value=0.5 \
    --result-file="data/results/kshap/kshap_2000s_100i.hdf5"

XGrad-CAM/Grad-CAM++

Example command: Run XGrad-CAM for the first 100 COCO images, for thresholding value of 0.5, and store results.

python scripts/gradcam_ssd.py \
    --first-images=100 \
    --explainer-engine="XGrad-CAM" \
    --baseline-value=0.5 \
    --result-file="data/results/cam/xgradcam_100i.hdf5"

Example command: Run Grad-CAM++ for the first 100 COCO images, for thresholding value of 0.5, and store results.

python scripts/gradcam_ssd.py \
    --first-images=100 \
    --explainer-engine="Grad-CAM++" \
    --baseline-value=0.5 \
    --result-file="data/results/cam/gradcampp_100i.hdf5"

Results

Kernel SHAP

• Kernel SHAP trains a linear explanation model to locally approximate the true model (using LIME framework)
• Exact runtime grows exponentially with the number of features, i.e. pixels.
• To constraint the runtime, we
  • segment image into 256 SLIC superpixels, each corresponding to one feature
  • train a surrogate model on 2000 samples (perturbations) per explanation (detected box)
• Evaluated on the first 100 images of COCO 2017 using SSD detector

Examples:

Correct detection Incorrect detection

Pixel flipping

• Area Under the Curve (trapezoidal rule): ~ 7.023 (unnormalized), or 0.0236 (normalized against 254 boxes)

Localization metrics

• Positive attribution inside box (against all positive attribution): 23.6%
• Box coverage: 64.8%
• IoU: 20.9%

Runtime

• On Google Colab's Tesla P100
• Batch size = 128
• 100 images (254 boxes): 2 hours 6 minutes
• Average time per explanation (i.e. box): ~30 seconds

XGrad-CAM/Grad-CAM++

• Implement two Class Activation Mapping (CAM) methods: XGrad-CAM and Grad-CAM++
• These two methods are the generalized version of Grad-CAM
• Grad-CAM is a model-specific local XAI method that extracts interested CAM visualization by weighting the feature map at one particular convolutional layer with its gradient caused by backpropagating a target prediction
• Runtime grows exponentially with the model complexity and linearly with the number of pixels

Examples:

• XGrad-CAM Correct detection Incorrect detection
• Grad-CAM++ Correct detection Incorrect detection

Pixel flipping

• XGrad-CAM: Area Under the Curve (trapezoidal rule): ~ 9.457 (unnormalized), or 0.0372 (normalized against 254 boxes)
• Grad-CAM++: Area Under the Curve (trapezoidal rule): ~ 13.024 (unnormalized), or 0.0512 (normalized against 254 boxes)

Localization metrics

XGrad-CAM:

• Positive attribution inside box (against all positive attribution): 18.65%
• Box coverage: 64.74%
• IoU: 16.93% Grad-CAM++:
• Positive attribution inside box (against all positive attribution): 14.59%
• Box coverage: 72.96%
• IoU: 13.84%

Runtime

• On Intel(R) Core(TM) i5-7360U CPU @ 2.30GHz
• 100 images (254 boxes): 38 minutes
• Average time per explanation (i.e. box): ~9 seconds

Pixel flipping comparisons

• Random flip (seed=0): Area Under the Curve (trapezoidal rule): ~ 82.188 (unnormalized), or 0.324 (normalized against 254 boxes)
• Summary plot
• All three methods yield on point explanations for detected objects
• kSHAP is slightly better than the two Grad-CAM methods
• Among Grad-CAM methods, XGrad-CAM is slightly better than Grad-CAM++