Research map / Architectures and efficiency / Pruning and efficient tuning

Temporal Action Detection Model Compression by Progressive Block Drop

CVPR 2025 · Architectures and efficiency › Pruning and efficient tuning · approach: Joint pruning and quantization · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Compresses temporal action detection models by progressively dropping redundant transformer blocks in the video feature extractor, recovering accuracy with LoRA-based cross-depth and prediction-level alignment to a frozen uncompressed teacher.

Problem

Temporal action detection models are computationally expensive, mainly due to the segment-level video feature extractor, and channel pruning yields small matrices that run inefficiently on GPUs.

Approach

Reduce network depth rather than width: iteratively drop the block whose removal least changes the training mAP, then recover performance with LoRA in attention layers trained to align intermediate features (cross-depth alignment) and predictions (KL and GIoU) with the uncompressed model.

Key results

• On THUMOS14 with VideoMAE-S, dropping 2 blocks gives 71.37 mAP vs 70.43 baseline at 84.1% MACs; dropping 4 blocks gives 69.65 mAP at 68.2% MACs.
• On ActivityNet-1.3 with VideoMAE-S, dropping 1 block gives 37.94 avg mAP vs 37.75 baseline at 91.7% MACs.
• Combining with channel pruning reaches about 30% MACs reduction at nearly identical accuracy (70.22 vs 70.21 mAP), and block drop gives 1.29x speedup vs 1.09x for pruning at similar MACs.

Measured effects reported in the paper

Surprising findings

• Pruning blocks progressively can improve mAP over the uncompressed model despite ~25% fewer MACs.
• Directly supervising the pruned model with ground-truth labels alone failed to restore performance, motivating feature-level alignment.

Limitations

• Iterative procedure requires evaluating every candidate block drop on the training set each step, which adds cost.
• Gains from block dropping diminish or reverse when too many blocks are removed (e.g., 4 blocks on THUMOS14 falls below baseline; 3 on FineAction degrades).

Experimental setup

Datasets: THUMOS14 ActivityNet-1.3 FineAction Charades-STA Kinetics-400

Benchmarks: THUMOS14 ActivityNet-1.3 FineAction Charades-STA

Metrics: mAP MACs inference time R1@0.3 R1@0.5

Baselines: channel pruning (Ding et al. structured pruning) uncompressed VideoMAE-S + ActionFormer random block drop AdaTAD ActionFormer with I3D

Builds on: ActionFormer AdaTAD LoRA VideoMAE curriculum learning

Closest papers by meaning

• End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames (CVPR 2024 · similarity 0.86)
• Dual DETRs for Multi-Label Temporal Action Detection (CVPR 2024 · similarity 0.86)
• Benchmarking the Robustness of Temporal Action Detection Models Against Temporal Corruptions (CVPR 2024 · similarity 0.86)
• Adapting Short-Term Transformers for Action Detection in Untrimmed Videos (CVPR 2024 · similarity 0.85)
• TE-TAD: Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression (CVPR 2024 · similarity 0.85)
• DiGIT: Multi-Dilated Gated Encoder and Central-Adjacent Region Integrated Decoder for Temporal Action Detection Transformer (CVPR 2025 · similarity 0.84)