1. Feature
I have tried both the sift and hog feature implementation. HoG is faster when the amount of crops are small, but as I reduce the parameters in the run_detector.m, d_sift becomes faster than hog in detection time. I think it's probably because that I wrote the wrapper function for sift but for hog I have to call the crops2feature in run_detection. All the results in this report are implemented by sift feature.
2. Mining hard negative
My get_hard_negatives function is very similar to the run_detection. The max_crops_per_scn is 100. For each image, I will sort the detections in the decending order of confidences and then send them to detection2bbox.m and bbox2crops.m. If the size of crops exceeds the max number of crops per image, I will take only the first max_crops_per_scn images(since they have the highest confidence). In the end, if the size of final crops exceeds the num_crops, I will random sample num_crops crops from the final crops.
With lambda =100, step_size = 2, scale_factor = 1.2, start_scale = 1,num_negative_examples = 1000. My linear svm with only random negatives can get 0.65 ap vs 0.774 ap by mining hard negative version of linear svm(total_stages = 3).
3. linear vs nonlinear
The linear svm is not very sensitive to the change of lambda, as long as it stays in a reasonable range. Results showed here are using lambda 100.
For nonlinear svm, my best rbf kernel parameters are lambda = 1.0, sigma = 256. The non-linear svm is quite sensitive to the variation of parameter.
With lambda =100 linear/1 nonlinear, step_size = 2, scale_factor = 1.2, start_scale = 1,num_negative_examples = 1000. My non-linear svm can get 0.864 ap vs 0.774 by linear.
With 5000 training examples, linear svm can get 0.784, but non-linear svm will run out of memory in training time.
Below are the Ap graphs for the above results
ap plot for linear svm with random negatives(1000 training examples).
ap plot for linear svm with mining hard negatives.
ap plot for linear svm with mining hard negatives and 5000 traning exmaples.
ap plot for the non-linear svm with random negatives
ap plot for the non-linear svm with mining hard negatives
login: wchi